ABSTRACT
The restoration of degraded ecosystems is critical for mitigating climate change and reversing biodiversity loss. Depending on the primary objective—such as maximizing carbon sequestration or protecting threatened species—and within the boundaries of budget limitations, different spatial priorities have been identified at global and regional scales. Funding mechanisms to support such work comprise public sources, philanthropy, and the private sector, including the sale of carbon and biodiversity credits. However, effectively exploring trade‐offs among restoration objectives and estimating the price of biodiversity and carbon credits to design financially viable projects remains scientifically challenging. Here we harness the power of artificial intelligence implemented in the software CAPTAIN to identify spatial priorities for ecological restoration that maximize multiple objectives at once and to allow a robust evaluation of biodiversity and climate outcomes. We find through a series of simulations that even low‐to‐moderate consideration of biodiversity in the design of restoration projects leads to the selection of restored areas that disproportionately improve the conservation of threatened species, while the carbon captured is only slightly reduced. We propose the bundling of biodiversity and carbon credits to support restoration efforts even in areas previously excluded for economic reasons. Applying our methodology to plant diversity in the Atlantic Forest of eastern South America, one of the most biodiverse and threatened ecosystems globally, we demonstrate its potential to guide restoration and to quantify trade‐offs between multiple optimization scenarios. Our study provides a robust, scalable method to maximize the joint benefits of ecological restoration for climate and nature.
Keywords: biodiversity financing, CAPTAIN, carbon sequestration, extinction risk, nature credits, reinforcement learning, systematic conservation planning
We present a reinforcement learning model for multi‐objective spatial restoration that jointly optimizes biodiversity, carbon sequestration, and costs. Applied to the Atlantic Forest, minor biodiversity weights yield disproportionate extinction risk reductions with small carbon loss. Bundling carbon and biodiversity credits can provide crucial support for financially viable restoration programs.

RESUMO
A restauração de ecossistemas degradados é fundamental para mitigar as mudanças climáticas e reverter a perda de biodiversidade. Dependendo do objetivo principal—como maximizar o sequestro de carbono ou proteger espécies ameaçadas—e dentro dos limites impostos por restrições orçamentárias, diferentes prioridades espaciais têm sido identificadas em escalas globais e regionais. Os mecanismos de financiamento para apoiar esse tipo de trabalho incluem fontes públicas, filantropia e o setor privado, incluindo a venda de créditos de carbono e de biodiversidade. No entanto, um grande desafio científico é como explorar de forma eficaz as perdas e ganhos entre objetivos alternativos de restauração, e como estimar o preço de créditos de biodiversidade e de carbono para o planejamento de projetos financeiramente viáveis.
Neste artigo, utilizamos o poder da inteligência artificial implementada no programa CAPTAIN para identificar prioridades espaciais para a restauração ecológica que maximizem múltiplos objetivos simultaneamente, e para permitir uma avaliação robusta dos resultados para a biodiversidade e o clima. Por meio de uma série de simulações, constatamos que mesmo uma consideração baixa a moderada da biodiversidade no planejamento de projetos de restauração leva à seleção de áreas restauradas que aumentam de forma desproporcional a conservação de espécies ameaçadas, enquanto o carbono capturado é reduzido apenas ligeiramente. Propomos a combinação de créditos de biodiversidade e de carbono para apoiar esforços de restauração mesmo em áreas anteriormente excluídas por razões financeiras.
Ao aplicar nossa metodologia à diversidade de plantas da Mata Atlântica—um dos ecossistemas mais biodiversos e ameaçados do planeta—demonstramos o potencial da nossa ferramenta para planejar a restauração sob múltiplos cenários de otimização. Nosso estudo fornece um método robusto e escalável para maximizar os benefícios conjuntos da restauração ecológica para o clima e a biodiversidade.
1. Introduction
Ecological restoration is one of the most powerful nature‐based solutions to tackle the twin planetary crises of biodiversity loss and climate change (Seddon et al. 2020). In recognition of its importance, the United Nations has declared 2021–2030 the Decade on Ecosystem Restoration and it motivates the commitment under Target 2 of the Kunming‐Montreal Global Biodiversity Framework—adopted by 196 countries—that “by 2030 at least 30 per cent of areas of degraded terrestrial, inland water, and marine and coastal ecosystems are under effective restoration”. In addition to protecting natural ecosystems—the preferred option whenever possible—well‐designed and implemented ecological restoration can capture and store carbon, increase populations of threatened species and reinstate or strengthen many of nature's contributions to people, helping us achieve several of the Sustainable Development Goals (Weltbiodiversitätsrat IPBES 2019; Di Sacco et al. 2021; Brancalion et al. 2025).
Despite the multiple benefits of restoration and its urgency in the face of global environmental change, two major challenges hamper its implementation. Firstly, much like prioritizing areas for protection (Jung et al. 2021), there are clear trade‐offs when designing restoration projects aiming to maximize carbon capture or biodiversity recovery (Strassburg et al. 2020). For example, the planting of exotic trees across Africa, although they grow quickly and capture large amounts of carbon (in principle helping to mitigate climate change, if the biomass produced is retained long‐term), has led to detrimental effects on native biodiversity (Parr et al. 2024). Trade‐off effects have been widely discussed in the literature [e.g., (Holl and Brancalion 2020; Hua et al. 2022)], but to our knowledge no tools have yet been developed to allow their quantification while taking into account the complexity of the real world—including the spatial heterogeneity in biodiversity, carbon storage potential, costs, and their changes over time.
Secondly, there is a major funding gap—estimated to exceed 600 billion USD annually (Deutz et al. 2020)—to implement the Global Biodiversity Framework, including its restoration target. Given that public funding from institutions such as the Global Environment Facility, which provides funding of roughly USD 1.3 billion per year, falls far short of reaching the required levels, many restoration projects today are supported instead by the sales of carbon credits, which can be voluntary or mandatory depending on the country. Alas, the simplicity of carbon credits, where one credit equals one tonne of atmospheric carbon captured, regardless of location, is also a disadvantage from the viewpoint of biodiversity. As market forces push trade towards low prices, they incentivize restoration projects that minimize costs (such as reforestation on marginal and highly degraded land) and focus exclusively on maximum carbon recovery, even at the cost of biodiversity benefits. To improve this situation, carbon credits are being designed to include additional benefits such as biodiversity, and in parallel, biodiversity credits are emerging following strict principles for high integrity and governance (Antonelli et al. 2024; Maron et al. 2025). Quantifying the cost relationship between carbon and biodiversity credits to make them financially interchangeable (allowing, for instance, a restoration project to become viable by selling biodiversity credits rather than carbon credits) can create economic opportunities for bundling carbon and biodiversity credits for the same site, while avoiding double claims. This strategy can enable the restoration of ecosystems in expensive areas, which have so far been excluded due to insufficient profit margins (Torabi and Bekessy 2015).
Here, we propose an approach based on multi‐objective reinforcement learning to address these shortcomings. Specifically, we: (i) show through simulations that this approach can help spatially guide restoration programs while monitoring their outcomes over time; (ii) explore trade‐offs and synergies that exist between restoration efforts focused on carbon, biodiversity, and costs through multi‐objective optimization to identify socially optimal solutions; (iii) link our predictions to the emerging concept of biodiversity credits alongside carbon credits to help design financially viable restoration programs; and (iv) run our optimization framework as a proof‐of‐concept on a large empirical dataset from the Atlantic Forest in South America. Our results show that identifying trade‐offs between carbon and biodiversity is crucial to evaluate the potential impact of biodiversity credits and to achieve restoration programs that maximize benefits for both climate and biodiversity.
2. Methods
2.1. Reinforcement Learning Optimization of Restoration
We optimized spatial restoration planning by developing new functionalities within the CAPTAIN software (Silvestro et al. 2022). We used environments and species generated through simulations to assess the performance and results of different restoration models under fully controlled settings. However, the framework can be used with empirical data, for instance by using species distribution models, climate data, and land use and other indexes of human footprint, and real costs of restoration. Similarly, empirical data, when available, can be used to inform the model about species dispersal ability, sensitivity to disturbances, and—in the case of plants—net sequestered carbon. In the following sections we first present how the updated CAPTAIN framework works for restoration with multi‐objective reinforcement learning and then provide the details of how we set up the specific experiments presented in this study.
2.1.1. Environment
The environment is discretized into spatial units of arbitrary size that represent the cells of a grid and are used as units for the description of environmental variables, species ranges, and restoration actions. Each cell features environmental variables distinguishing land from sea and providing elevation and climate data (such as temperature and precipitation—for example, their mean, monthly and seasonal variations; for ocean environments, specific variables such as water depth, oxygen levels, and salinity could be used). Cells are also characterized by a value of anthropogenic disturbance, indicated with , that alters the ability of species to live in the cell and their abundance (see below) and by a restoration cost (), quantifying the relative cost of restoring the cell to a low level of disturbance. This could include, for instance, the cost of buying the cell, displacing or removing the cause of disturbance, and potentially taking action to reintroduce native species. Note that in real conservation work, restoration costs may be complex depending on the legal and socio‐economic setting. For example, buying land is not always sufficient to stop all logging, but we assume that we can model appropriate total costs here. All these features can be set to change through time, for example, capturing the dynamics of predicted climate change, variation in disturbance patterns, and evolving costs per cell.
The environment is also characterized by the species that live (or can live) in it. Specifically, for each cell and each species in the system, there is a value describing the number of individuals that can occupy the cell in the absence of disturbance (i.e., the natural state). This is a cell‐ and species‐specific carrying capacity which depends on both the habitat suitability of the cell for that species () and a species‐specific maximum abundance per cell ():
| (1) |
The environment also includes the actual number of individuals of a species at any given time , , which results from the interaction between the species and the environment, including the effects of disturbance, growth rates, and dispersal.
For each species, we define a specific value of sensitivity that determines individual mortality fractions in response to a disturbance in a cell, . Species are also assigned a growth rate , which describes the demographic growth in one time step, and a dispersal rate , which determines the number of individuals that disperse from one cell to another. In the current implementation, we assume individuals to be static (e.g., plants). Therefore, dispersal does not move individuals from one cell to another. Instead, dispersal only adds individuals to the receiving cell (replicating for instance the dispersal of seeds). The number of individuals dispersing from cell to follows a truncated exponential distribution:
![]() |
(2) |
where is the distance between the two cells and is the maximum distance that can be reached in a single time step. Species distributions and abundances change over time based on growth, dispersal, and mortality. Specifically, at each time step, the distribution and abundance of species are updated according to the following equation:
| (3) |
where is the number of individuals of species in cell at the previous time step, is the product between the disturbance in the cell at time t and the species sensitivity, with representing the mortality driven by disturbance. The term represents the contribution of demographic growth, while
| (4) |
is the sum of the individuals dispersing from all other cells and represents the contribution of dispersal to the cell's population.
Each species in the environment is assigned a value representing its expected carbon value . This is a single value and can be interpreted as the net amount of carbon sequestered by one individual of species . As our simulated system in its current implementation does not track the age of each individual, we define the total carbon in a cell as the sum of carbon storage from all individuals and all species in the cell:
| (5) |
This quantity could in principle be extended to include carbon in the soil. We note that, as the age of individuals is not modeled here, the simulated accumulation of carbon in the current implementation might occur more rapidly than in empirical systems.
Finally, we also include a 5‐tiered classification of species extinction risk, inspired by the International Union for Conservation of Nature (IUCN) (IUCN 2023). When working on empirical data, this can be effectively replaced by actual IUCN Red List assessments or model‐based predictions (Zizka et al. 2022; Antonelli et al. 2023). Hereafter, we will use the Red List‐inspired labels to describe increasing extinction risks: Least Concern (LC), Near Threatened (NT), Vulnerable (VU), Endangered (EN), and Critically Endangered or (locally) extinct (CR). As part of the evaluation of the implemented restoration action, the initial extinction risk of a species can be updated based on the simulated dynamics of the species demography or geographic range, for instance, increasing or decreasing the extinction risk depending on whether the census population size of a species is decreasing or increasing, respectively. We indicate the extinction risk of each species at time as:
2.1.2. Agent
The agent in the system observes the environment at each time step extracting features for each grid cell about its disturbance level, its cost for restoration, the species that occur there, those that could occur there naturally in the absence of disturbance, their current extinction risk and carbon contribution (a detailed list of the features used in our experiments is provided in the next section). Following the modeling framework developed in the initial implementation of CAPTAIN (Silvestro et al. 2022), the features of each cell feed into a fully connected neural network with a tanh activation function (Apicella et al. 2021) in the hidden layers and a single output node. The parameters of the neural networks are shared across all cells and their outputs are pooled into a single vector of values providing a restoration rank of each cell. The agent will then select deterministically a predefined number of cells based on their ranking, within which restoration occurs. Here we assume restoration to result in the instantaneous removal of disturbance from the cells, although different dynamics such as a gradual decrease or partial reduction in disturbance can be implemented. The restoration action is associated with an economic cost computed as the cumulative cost of the selected cells. Once the disturbance is removed from a cell, its populations will continue to evolve according to Equation (3) with reduced (or removed) disturbance.
2.1.3. Rewards and Optimization
The actions of the agent result in changes in the environment that are quantified through a total reward, defined based on the objective of the restoration policy. The parameters of the neural networks determining the restoration action are chosen to maximize the cumulative net reward at the end of a simulation of a fixed number of steps. The net reward or payoff is here defined as the weighted sum of three components quantifying net carbon gain (carbon reward), net biodiversity outcomes (biodiversity reward), and costs (budget reward). These quantities allow us to optimize policies with one or multiple objectives, generating socially optimal solutions when multiple objectives are used. At each time step, the carbon reward of a restoration action is computed as the ratio between the net change in total carbon stored in the system compared with its initial state (at time ) and the amount of carbon lost between the natural state (in the absence of any human disturbance) and the initial state:
| (6) |
We note that with simulated datasets, like the ones used in this study, the natural carbon amount , is known a priori. When working with empirical data, this quantity might have to be approximated based on measurements, such as through survey data (for trees, this can be derived indirectly from measurements of height and stem diameter at breast height (dbh), or more directly from biovolumes estimated from lidar scans).
The total reward of a restoration action also includes a measurement of the resulting biodiversity outcome, which can be defined in many different ways, as shown by a recent compendium (Burgess et al. 2024) currently listing over 570 biodiversity metrics. In the current implementation, we exemplify the calculation of biodiversity gains in a similar way as recently calculated in (Silvestro, Goria, Groom, et al. 2025), where the number of extinct or threatened species was used as a negative reward (see also (Possingham et al. 2015)). The focus on extinction risk is also aligned with other recently proposed metrics for conservation priorities and threat abatement (Durán et al. 2020; Mair et al. 2021; Eyres et al. 2024). Here we use an average of the restored species population weighted by their current conservation status. Specifically, for any given extinction risk status , we identify the subset of species such that:
| (7) |
and compute the respective reward as
| (8) |
where is a weight associated with extinction risk status , and is the ratio between the number of individuals of species at time found within restored cells (indicated with ) and the total population of the species in the environment:
| (9) |
Thus, when the species is not found in any restored areas and when all its population falls inside restored areas. The weights are set to favor scenarios where the highest average fraction of population is restored in species classified in a high‐risk category:
| (10) |
so that, for example, protecting 30% of the individuals in a Least Concern species () contributes 0.0025 to the reward, while the same fraction in a Critically Endangered species () contributes 0.1587. The overall biodiversity reward is set to the sum of over all extinction risk categories. The use of increasing weights associated with extinction risk is consistent with several metrics developed to quantify conservation priorities (Mair et al. 2021; Pimiento et al. 2020; Gumbs et al. 2023).
Finally, we include the restoration costs as part of the reward. Based on an arbitrarily large reference budget we compute the budget reward as
| (11) |
where is the total restoration cost summed over all restored cells (defined in the next section).
The rewards can be used alone or in combination to optimize policies that pursue one or more targets. Thus, we defined the total reward as:
| (12) |
where the predefined weights () quantify the proportional importance assigned to carbon, biodiversity, and budget. Here we set the weight associated with the budget () to a constant and set the others to express their relative importance. Under this formulation, a policy targeting a single objective (e.g., focusing exclusively on budget) can be optimized by setting the respective weight to 1 and the others to 0 (e.g., , and ). The weights can also be used to estimate different solutions where the total reward includes two or more potentially competing optimality criteria.
The CAPTAIN optimization algorithm trains a neural network policy to maximize the expected reward from protective actions by implementing a hybrid approach that combines parallelized evolution strategies (a genetic strategy algorithm) with concepts from classical policy gradient methods, that is, an advantage function (Silvestro et al. 2022; Sutton et al. 1999; Salimans et al. 2017). The algorithm operates through a repeated two‐phase process (epochs): first, the policy parameters are randomly perturbed and evaluated via parallel environment simulation episodes; second, the results from these parallel runs are aggregated, and the parameters are updated using a stochastic gradient estimate. The parameter update is weighted by the advantage function, which measures the improvement of the current average reward relative to the running reward (weighted average of rewards across different epochs) (Silvestro et al. 2022). This approach allows the policy to be optimized without the computational burden of calculating explicit gradients while enabling parallelization of the optimization task.
2.2. Simulations
We use simulations of biodiversity to evaluate the ability of restoration policies optimized through reinforcement learning to achieve different targets, focusing on carbon and biodiversity outcomes. Although the same policies could be applied to empirical data, simulations allow us to benchmark and compare different restoration actions across many simulated scenarios in which we can accurately measure all outcomes.
2.2.1. Parameters of the Simulated Environment
We simulated environments based on a raster of 128 × 128 cells, of 25 km resolution at the equator, inspired from physical properties of an arbitrary region spanning a portion of Central and South America (Figure S2a). The properties include its coastline, elevation and cell‐specific climate average values obtained from CHELSA (Karger et al. 2023), here including the mean annual temperature, mean annual precipitation and the minimum temperature of the coldest month. We used the gdgtm Python library to generate the initial rasters (see Supplementary Data and Codes). The selection of variables is not intended to provide an exhaustive account of the climate variables that determine a species' niche but was used to generate artificial species distributions with realistically constrained geographic ranges. After ‐transforming these topographic and climate variables we indicate them as , where represents the value of the variable corresponding to the longitude and latitude coordinates .
We simulated species geographic range by first drawing a random pair of coordinates (), based on which we extracted the corresponding elevation and climate data. We use a normal distribution to describe the environmental and climatic niche of the species , where the diagonal of the covariance matrix, a vector , is randomly drawn for each species from a uniform distribution and off‐diagonal values (covariances) are set to 0. The vector will therefore generate species with a range of narrow to wide climatic tolerances. We also introduce a distance factor, to ensure that species ranges are also spatially constrained, for example, through limited dispersal ability and not only determined by habitat suitability. The effect of distance was simulated through an exponential distribution with rate parameter randomly drawn for each species from a uniform distribution . Thus, for any set of coordinates () the relative population density of a species is , where is the Cartesian distance between and . We use this formula to define the habitat suitability of a species in each cell , as used in Equation (1).
We used the empirical mean annual precipitation layer (rescaled between 0 and 1) to split cells into three categories with high (> 0.4), intermediate (0.25–0.4) and low precipitation (< 0.25). We use these categories to simulate different vegetation types, for example, representing one dominated by trees, a mixed open habitat, and a fully open habitat (Figure S2a). These vegetation types are then used to characterize properties of the species that inhabit them. Specifically, we simulated three categories of species to represent different growth forms, for example, herbaceous plants, shrubs, and trees. We assigned each species to one of the categories randomly with probability set equal to the proportion of the overlap of their simulated range with low, intermediate, and high precipitation, respectively.
We drew the species demographic growth rates from a Weibull distribution:
where 10, 2, and 0.1 for the first, second, and third growth forms, respectively. This results in faster average demographic growth for simulated “herbaceous” species. Similarly, we drew the species‐specific carbon value from a gamma distribution:
where 0.5, 1, or 2 is the mean of the distribution applied to the first, second, and third growth forms, respectively. We draw the species‐specific carrying capacity (i.e., the maximum number of individuals that can live in a cell; Equation (1)) from a beta distribution that generates many species with low abundance and few species with high abundance:
This leads to a mean carrying capacity of a species per cell of ≈35 ranging from ≈2 to ≈102. We assigned these values by setting the largest ones to species of the first growth form, the intermediate values to those of the second, and the smallest values to species of the third. Thus, species with fast growth and low carbon value tended to have higher carrying capacity than slow‐growing species with high carbon value (reflecting the realistic differences in wood densities and carbon storage among woody species). Examples of the simulated species' natural abundances resulting from both habitat suitability and species‐specific carrying capacity are shown in Figure S2d.
2.2.2. Disturbance and Restoration Costs
We simulated spatially heterogeneous disturbance through random multivariate normal distributions rescaled to range between 0 and 1, which we multiplied by exp(‐rescaled elevation) to simulate a scenario in which anthropogenic disturbance decreases on average with elevation (given the increased difficulties in accessing natural resources or performing agriculture; see example of simulated disturbance in Figure S2c). We assigned a sensitivity to disturbance for each species based on a symmetric U‐shaped beta distribution:
thus generating several species with either very low or very high sensitivity. Prior to applying any restoration policy, the environment was set to evolve for a number of steps until reaching an equilibrium under the simulated disturbance.
The restoration costs were linked to disturbance, so that on average the cost of restoration was higher for highly degraded cells (reflecting fewer opportunities for natural regeneration and higher distance to seed sources). We added some stochasticity in this pattern by computing costs as the product between disturbance and another rescaled random multivariate normal distribution (see example of simulated disturbance in Figure S2c). We set the top 5% of cells with the highest cost as unrestorable, removing them from the set of cells that can be restored with an action from the agent. These cells are meant to represent areas where restoration is not possible, for example, because that would imply displacing inhabited areas or activities and infrastructure that cannot be moved.
2.2.3. Evolution of the Species Extinction Risk
We initialized the species extinction risk based on their range size, on their sensitivity, and on the effect of disturbance:
| (13) |
where is the natural range of the species (number of occupied cells) in the absence of disturbance. Thus, the value decreases with small natural range sizes and with increasing sensitivity and/or disturbance (Strassburg et al. 2020). The exponent (here set to ) has the effect of flattening the difference in initial extinction risk based on natural range size. We then mapped the value to an initial extinction risk (Table 1). We note that this approach to simulate species extinction risk, while based on common assumptions about species range size and decline due to anthropogenic disturbance, is arbitrary and should be replaced with actual risk assessments (e.g., from the IUCN Red List) when running the optimization with real world data.
TABLE 1.
Thresholds used to set the initial species extinction risks (, only applied in simulated data and replaced with Red List assessment when using empirical data) and their reassessment at time based on their level of protection (Criterion 1) and population trends (Criterion 2). The re‐assessed extinction risk is .
| Initial status | Reassessment at time | ||||||||
|---|---|---|---|---|---|---|---|---|---|
|
|
|
Criterion 1 |
|
Criterion 2 |
|
||||
| LC | > 0.95 |
|
< 0.10 |
|
< 0.10 | ||||
| NT | 0.9–0.95 |
|
0.10–0.25 |
|
0.10–0.25 | ||||
| VU | 0.8–0.9 |
|
0.25–0.50 |
|
0.25–0.50 | ||||
| EN | 0.6–0.8 |
|
0.50–0.80 |
|
0.50–0.80 | ||||
| CR | < 0.6 |
|
> 0.80 |
|
> 0.80 | ||||
Throughout the simulations, the extinction risks of all species were re‐evaluated at each step, and we used two criteria to re‐assess the conservation status of a species. First, we calculated the number of restored cells within which a species occurred at time , indicated as relative to the species natural range size . Thus, increasing values of indicate that a proportionally higher number of restored cells is occupied (potentially after re‐colonization) by species , leading to a decreased extinction risk (Table 1, Criterion 1). Second, we calculated the ratio between the occupied geographic range at time and the occupied range at the beginning of the simulation , that is, prior to enforcing restoration programs. This ratio is used to verify if the species is in decline, for example, if the restoration is ineffective or if other factors, such as climate change, negatively affect a species' range. We used the thresholds reported in Table 1 (Criterion 2) to determine if the species was in decline and to what extent that affected the species extinction risk. Finally, any species among the initial simulated pool with occupied range size of 0 was moved to the highest risk category and considered as possibly extinct in the study system. These criteria are inspired by some of the Red List guidelines (IUCN Standards and Petitions Committee 2025), specifically looking at species range sizes, and recovery or declining trends through time. While we used the occupied range as the main metric, the criteria can be defined based on alternative metrics, for example, based on census population size and its temporal variation or fragmentation of the species range.
2.2.4. Observed Features
The agent was set to observe the entire environment at each step, extracting several features for each cell, which form the input of the neural network and determine the next action. We used the same 16 features for all policies, independently of the reward. The first features were the disturbance in the cell and the mean disturbance in the 5 × 5 cells around it (features 1–2). The latter feature provides a measure of the spatial distribution of disturbance, for instance, differentiating between a disturbed cell neighboring a pristine area and one placed within a broader degraded area. We then included the potential carbon net change in the cell calculated as the difference between the total carbon in the cell under natural conditions and the current amount of carbon, that is, . The same measure is also provided as a mean value in the 5 × 5 cells around the cell (features 3–4). The features additionally include the number of CR, EN, VU, and NT species in the cell based on the current populations occupying the cells and the mean number of CR and EN species in the surrounding 5 × 5 cells (features 5–10). We also included the number of species in each of the five extinction risk categories, but counting only species that do not occur in any of the previously restored cells, if any (features 11–15). These features serve as a measure of complementarity among cells so that the model can choose to restore cells with a different species composition compared with areas already restored. Finally, we included the cell's cost (feature 16). All features were rescaled between 0 and 1 before being fed to the neural network.
2.2.5. Training
We trained models with 250 species running simulations for 30 time steps (such as years). The agent was set to restore 100 cells at each step, reaching 3000 restored cells at the end of the simulation, that is, corresponding to about 30% of the land cells in our grid to mimic the 30% restoration target of the Kunming‐Montreal Global Biodiversity Framework. We used a parallel genetic strategies algorithm with 8 environments evolving in parallel and ran 5000 epochs to train the models. We monitored the reward across epochs to assess model convergence (Figure S4). We used small neural networks for each cell, connecting the 16 features to a first hidden layer of 3 nodes, followed by a second hidden layer of 2 nodes. The parameters of the network were shared among all cells. We also tested a larger neural network model, which however did not lead to significant performance improvements in our experiments (see Appendix S1; Figure S5). We trained different models exploring a gradient of optimality (Lin et al. 2019) going from favoring the carbon outcome to favoring the biodiversity outcome. We achieved this by assigning different weights to the reward (see Equation (12)). Specifically, we trained five models, keeping the weight of the budget reward constant at 0.5 and assigning different weights to carbon and biodiversity such that the sum of the weights was kept constant and equal to 1 (Table 2).
TABLE 2.
Weights used in the different trained models and assigned to the carbon reward (), biodiversity reward (), and budget reward ().
|
|
|
|
||||
|---|---|---|---|---|---|---|
| Policy 1 | 0.5 | 0 | 0.5 | |||
| Policy 2 | 0.4 | 0.1 | 0.5 | |||
| Policy 3 | 0.25 | 0.25 | 0.5 | |||
| Policy 4 | 0.1 | 0.4 | 0.5 | |||
| Policy 5 | 0 | 0.5 | 0.5 |
Note: We considered a range of models keeping the weight on the budget reward constant, while progressively moving from a carbon‐focused optimization (Policy 1) to a biodiversity‐focused one (Policy 5). The intermediate policies allowed us to explore the trade‐offs between these potentially competing objectives.
2.2.6. Experiments on the Reference Region
After training the models, we used them on 100 simulations each including 500 species and initialized following the same approach utilized to generate the training data. These simulations were run on the same reference region used for the optimization of the policies. Although these simulations were based on the same simulation settings, they differed from each other, being based on random draws for all parameters. Thus, the distribution of species, disturbance, and costs and the traits associated with each species (carbon, sensitivity, and growth rates) varied randomly across simulations. This allowed us to evaluate the results of different models across a range of different parameter values. The simulated environments in their initial state, that is, prior to restoration, included on average 30 CR species (standard deviation, sd: 18), 87 EN (sd: 39), 156 VU (sd: 31), 139 NT (sd: 30), and 87 LC (sd: 56) species. As during training, the simulated datasets evolved for 30 time‐steps with the agent interacting with the environment based on each of the five trained policies and restoring 3000 cells. We evaluated the outcomes of the different policies by calculating the amount of carbon gained through restoration, the changes achieved in terms of number of species falling in each of the five extinction risk categories, the number of species occurring within restored areas at the end of the simulation, and the total cost of restoration. This allowed us to quantify the trade‐off between carbon and biodiversity outcomes along a gradient of optimal solutions.
2.3. Carbon and Biodiversity Credits
We implemented a simplified concept of carbon credits by assigning a monetary value to the amount of net carbon gain (or potentially loss) incurred during the simulations. As carbon (like restoration costs and budget) is set in arbitrary units in our simulations, we defined the value of a carbon credit unit () as the value of the net carbon gain obtained through Policy 1, namely the policy placing the highest weight on carbon, leading to the full compensation of the restoration costs:
| (14) |
where is the value of one carbon credit unit based on the simulated systems used for carbon and budget, the fraction's numerator is the net carbon gain at time compared to the initial state () and the denominator is the total cost of restoration. Thus, under this definition, the carbon credits accumulated through the implementation of Policy 1 result, by construction, in full compensation of the cumulative restoration costs, with no economic loss at the end of the simulation. We note that this definition is arbitrary and could be replaced by other definitions, including empirical valuations of carbon in the cap‐and‐trade system of the European Union, or others.
The same carbon credit value was then applied to the other four policies where carbon was given lower weights (Policies 2–5) in the training phase with the inclusion of a positive weight assigned to the biodiversity reward. This allowed us to evaluate to what extent the carbon credits accumulated through restoration policies that optimized for both carbon and biodiversity (Policies 2–4) or biodiversity only (Policy 5) were able to compensate for the restoration costs.
Next, we selected a simple metric as a reference to define a biodiversity credit and developed an approach to value the credit relative to the value of a carbon credit. Specifically, we quantified the biodiversity credits as the number of species that moved from a threatened class (CR, EN, and VU) at the beginning of the simulation to a non‐threatened class (NT and LC) after restoration. This differs from current practice in some of the emerging biodiversity credit markets, such as in Colombia where one biodiversity credit unit equals 1 hectare regardless of how many species benefited from the restoration and in which ways. However, the same approach utilized here can be used for alternative definitions of biodiversity credits.
We then evaluated the net financial loss or gain of each policy after stacking the value of carbon and biodiversity credits under different valuations of biodiversity credits. We indicate with the value of one biodiversity credit obtained from moving a species from a threatened to a non‐threatened status after restoration. Given that we are using arbitrary units for costs and carbon credits, we defined the value of a biodiversity credit relative to that of a carbon credit. This allowed us to assess which policy would become financially viable or preferable, in our simulated context, if, for example, the value of one biodiversity credit equals that of a carbon credit (i.e., ). We tested valuations of and calculated the net financial loss or gain of a policy () as a percentage of the restoration cost:
| (15) |
where is the net carbon gain of the policy, is the policy's net biodiversity outcome (as defined above), and is the cost of restoration.
Finally, we calculated, for any given policy , the value that should be attributed to a biodiversity credit so that the policy results in the same net financial outcome as a restoration program focused on carbon (i.e., our Policy 1):
| (16) |
The value of one biodiversity credit is therefore, after solving Equation (16) for :
| (17) |
Using this formula, we derive the value of a biodiversity credit (relative to the value of a carbon credit) required to make the financial outcome (defined in Equation (15)) of any optimal restoration policy equal to the profit of a carbon‐oriented policy (our Policy 1), that is, such that . We note that under this definition, the value of biodiversity credits is context‐specific, as it depends on the overall area considered and its restorable carbon and biodiversity. Within the area, the restoration of cells that contribute the most to reducing the extinction risk of species (e.g., cells that protect rare and threatened species) will lead to the highest biodiversity credits (Drupp et al. 2024).
2.4. Empirical Analysis
We analyzed an extensive dataset of tree biodiversity in the Atlantic Forest (de Lima et al. 2024). The dataset included occurrence data for thousands of woody plant species in the Atlantic Forest, with per‐species information on their maximum height, wood density, and conservation status based on the IUCN Red List classification of extinction risks (IUCN 2023; de Lima et al. 2024).
We compiled spatial data gridded into 16,376 cells at a 10‐km resolution within the Atlantic Forest region (around 1.2 × 106 km2 in area) as defined by the WWF Global 200 project (Olson and Dinerstein 2002). We generated species distribution models (SDM) based on cell‐specific WorldClim bioclimatic variables (Fick and Hijmans 2017) and on species occurrence data (de Lima et al. 2024), and generated raster layers representing disturbance and restoration costs based on publicly available geospatial data products. All data processing and analyses were run in R 4.3.3, using the packages concaveman, ENMTools, flexsdm, future, fuzzyjoin, geodata, magrittr, sf, terra, tidyterra, and tidyverse (Pebesma 2018; Pebesma and Bivand 2023; Gombin et al. 2020; Warren and Dinnage 2025; Velazco et al. 2022; Bengtsson 2021; Hijmans et al. 2024; Bache and Wickham 2022; Hijmans 2025; Wickham et al. 2019; Hernangómez 2023; R Core Team 2024).
2.4.1. Species Distribution Models
We obtained the 19 WorldClim historical bioclimatic variables at 5 arc‐minutes resolution (Fick and Hijmans 2017), cropped to the Atlantic Forest region and reprojected to the 10‐km resolution. We examined the correlation between the bioclimatic variables to exclude highly collinear variable pairs, retaining the following 7 variables of temperature (°C) and precipitation (mm): (1) annual mean temperature, (2) mean diurnal temperature range (mean of monthly maximum temperature—minimum temperature), (3) annual temperature range, (4) annual precipitation, (5) seasonality of precipitation (coefficient of variation), (6) precipitation of the warmest quarter, and (7) precipitation of the coldest quarter.
Among the retained variables, only variable pairs 1–5 () and 2–7 () exhibited coefficients of correlation larger than 0.6 (Figure S10): as these are both pairs of a temperature variable and a precipitation variable and they may contain important environmental information despite the correlation, we retained both variable pairs.
We obtained geolocated occurrence data for 4950 species occurring in the Atlantic Forest (de Lima et al. 2024), and eliminated data points outside of the rectangular bounding box of the Atlantic Forest region to exclude geographically distant data points. The remaining data set contained in total 722,448 records and 4948 species. Of these, 3105 species had abundant data points () and 1631 species had sparse data points (); 212 species with insufficient data points () were excluded from the analysis. To mitigate potential spatial biases, we performed spatial thinning for species with abundant data points, using the ENMTools::trimdupes.by.raster function to retain one data point at maximum per pixel. To prevent excessive loss of species information, we only accepted the thinned results if they also contained data points; otherwise, we retained the original unthinned data. In total, 1843 species were thinned, and for the thinned species of data points were retained per species on average.
We defined the spatial extent of the training area (from which background points are sampled) for each species as an alpha shape (concavity = 3) surrounding all occurrence data points of the species, buffered by twice the median inter‐point distance among all data points and cropped to the Atlantic Forest region. We sampled 10,000 background points within the training extent of each species (Valavi et al. 2021), assembled the occurrence data points and the background points, and extracted the environmental values at all points to construct the model data. We partitioned model data into training and testing sets for cross‐validation, using complete random ‐fold partition implemented in the flexsdm::part_random function: we used 4‐fold partition for the 3105 species with abundant data points, and 2‐fold partition for the 1631 species with sparse data points.
We then fitted an SDM for each species: for the 3781 species with data points, we fitted a generalized additive model (GAM) using the flexsdm::fit_gam function, and for the 955 species with data points, we adopted an ensemble of small models (ESM) approach using the flexsdm::esm_gam function, which consists of fitting bivariate GAM models with all pair‐wise combinations of predictors, and calculating the ensemble average of model outputs weighted by their performance metrics (Somers' : ) (Breiner et al. 2015). For the model output of each species, we calculated the following five performance metrics: sensitivity, specificity, omission rate, true skill statistic (TSS), and area under curve (AUC). We excluded species with low performance (sensitivity , specificity , TSS , AUC ), retaining results for 3885 species (Figure S11). Finally, for the retained species, we generated geographical predictions of continuous suitability score maps over the Atlantic Forest region using the flexsdm::sdm_predict function.
2.4.2. Disturbance and Restoration Costs
We defined the disturbance level within each 10‐km pixel of the Atlantic Forest region based on its land cover. We obtained the MapBiomas Collection 10—Land Cover and Use dataset (MapBiomas 2024a), which classifies each 30‐m pixel within Brazil into one of 38 land cover classes. In addition, we also obtained the MapBiomas Collection 9—Age of Secondary Vegetation maps in 2023 (MapBiomas 2025), and created a mask of secondary forests representing pixels where the age of secondary vegetation . We cropped the datasets to the Atlantic Forest region, reprojected to the 10‐km resolution, and calculated the number of 30‐m pixels belonging to each land cover class within each 10‐km pixel, additionally classifying pixels within the secondary forest mask into a novel class of “secondary forest” (ID: 76). We reclassified the MapBiomas classes into the following 9 broad classes: Forest (IDs: 1, 3, 5, 6, 49); Grassland (4, 12, 29, 50); Secondary forest (76); Plantation (9); Cultivation (14, 15, 18, 19, 20, 21, 31, 35, 36, 39, 40, 41, 46, 47, 48, 62); Coastal area (23); Water bodies (11, 32, 26, 33); Urban area (24, 25, 30, 75); and Others. We calculated the proportion of each broad class within each 10‐km pixel. We then calculated the average disturbance score of each 10‐km pixel by assigning a disturbance score to each broad class, namely, forest and grassland: 0, secondary forest: 0.3, plantation: 0.5, cultivation: 0.7, coastal area: 0.5, water bodies: 1, urban area: 1, others: 0.5. We then calculated the weighted average of all disturbance scores weighted by the proportion of the broad classes.
We estimated restoration costs by integrating three spatial proxies: (1) accessibility (travel time to urban centers) and (2) elevation, which serve as proxies for logistical and implementation challenges, and (3) agricultural production value, representing the opportunity cost of land conversion. While this approach provides a robust approximation for our case study, we acknowledge that accurate restoration costs would require more granular data, including the costs of active reforestation, fencing, invasive species removal, and long‐term site management (Strassburg et al. 2020; MapBiomas 2024b; Adams et al. 2019). For accessibility, we retrieved data of travel time to cities with more than 5000 inhabitants (geodata::travel_time) (Brancalion et al. 2019). For elevation, we retrieved global elevation data (geodata::elevation_global) that are based on the Shuttle Radar Topography Mission (SRTM) 90‐m resolution data (Nelson et al. 2019) and supplemented by the GTOPO30 dataset for high latitudes (Farr and Kobrick 2000). For the value of agricultural production, we retrieved SPAM data of aggregated crop production value (USD) per area (geodata::crop_spam). SPAM pixels with empty values were filled in by interpolation from edges using the GDAL script gdal_fillnodata, with the md (maximum number of pixels to be searched for values to interpolate) set to 10. All three data layers were cropped and reprojected using the GDAL command gdalwarp, resampled to the 10‐km resolution with bilinear interpolation using the terra::resample function, and then scaled with min‐max scaling to the range of . We then calculated restoration costs as the arithmetic average of the three scaled variables.
2.4.3. Species Traits
We included species‐specific features in our analysis, allowing us to incorporate extinction risks, proxies for carbon sequestration potential, and life‐history traits. From the original data (de Lima et al. 2024) we obtained growth habit (categorized as shrub, small tree, and large tree), which we used to approximate relative values of growth rates. We used the same Weibull distribution used in the simulated data to draw species‐specific growth rates, with scale parameters set to 10, 2, and 0.1 for shrubs, small trees, and large trees, respectively. This led to faster average demographic growth rates for shrubs compared to large trees. We acknowledge that these are approximations and it would be preferable to replace this parameterization with empirical estimates of growth rates, where possible, for instance leveraging increasingly comprehensive databases such as TRY (U.S. Geological Survey 1996).
We randomly drew the values describing species‐specific sensitivity to disturbance from the same U‐shaped beta distribution used in our previous simulations. To simulate the fact that trees might be more sensitive to land use changes, we ranked the randomly drawn sensitivity values and assigned the highest values to large trees, followed by small trees, leaving the smallest values to shrubs. Again, these approximations should be replaced by empirical estimates of sensitivity to disturbance, where available. We expressed the carrying capacity of a species in a cell as a relative number between 0 and 100, scaled by the habitat suitability in the cell, as predicted by the species distribution models. The simulated abundance of a species in a cell was further affected by a mortality rate proportional to the product between disturbance and species sensitivity (Equation 3).
We used the available Red List assessments as species extinction risk estimation (IUCN 2023). Here we imputed missing extinction risk classification (i.e., unevaluated or data‐deficient species) through random resampling proportionally to the empirical frequencies of each class among assessed species. While this is a crude approximation, more sophisticated approaches could be used to predict extinction risks using machine learning imputation methods (Zizka et al. 2022; Kattge et al. 2011). We let the extinction risk of species evolve over time as a function of species recovery after restoration using the re‐assessment criteria used in our simulations (Table 1).
We used per‐species biomass as a proxy for a species' ability to sequester carbon, based on wood density (WD) and maximum height (H) (de Lima et al. 2024). We estimated the maximum diameter at breast height (DBH) for each species from maximum height (H), using the generic model 3 from (Estopinan et al. 2024): from which we obtain . Next we estimated maximum above‐ground biomass (AGB, kg) per tree for each species from wood density (WD), DBH and H, using the pantropical model from (Cysneiros et al. 2020): . As in the analysis of simulated data, we estimated the relative above‐ground carbon in a cell based on the species‐specific AGB estimated as shown above and the simulated abundance of the species in the cells. We then calibrated this relative quantification of carbon per cell, by rescaling it such that their average equaled 98.5 tonnes per hectare, based on previous empirical estimates for the Atlantic Forest (Chave et al. 2005). We then used the difference between the initial carbon and the total carbon after restoration as a measure of the sequestration generated by the restoration policy. While these are approximate estimates of potential carbon heterogeneities in the region based on species composition, advances in above‐ground carbon estimation for instance from remote sensing data (Bordin et al. 2023; Lamahewage et al. 2025) will likely improve the quality of the restoration prioritization in the future. For example, high‐resolution remote sensing data products (Zhang et al. 2025), in combination with ground data such as forest inventory plots (Lang et al. 2023) and terrestrial laser scanning (Lopez‐Gonzalez et al. 2011) in areas where they are available, could be used to provide more accurate estimates of the species‐specific carbon sequestration potential.
2.4.4. Model Fine‐Tuning and Predictions
We fine‐tuned the models by running another training procedure for 1100–1700 epochs starting from the parameter values from the models trained on simulated data (Policies 1, 3, and 5). The additional training was based on a batch of 6 parallel environments, each containing a random subset of 1000 species to speed up the run time.
We then used the fine‐tuned model on the full set of species to identify restoration priorities. As in the analyses of simulated data, we set the models to select 4800 grid cells (i.e., approximately 30%) within 30 time steps. Thus, at each time step 160 cells were selected for restoration, and the disturbance in protected cells was set to 0, allowing existing species in the cell to reach their carrying capacity and species from neighboring cells to recolonize them based on habitat suitability and a dispersal process modeled as in our previous simulations.
As some of the parameters in our simulations were imputed (namely, the extinction risks for unassessed species) or randomly drawn from prior distributions (i.e., the growth rates and sensitivity to disturbance), we repeated the restoration predictions 30 times, resampling each time these parameters. We then summarized the results by ranking cells based on how frequently they were selected for restoration and took the top 4800 cells as the final outcome. We additionally performed sensitivity analyses where we introduced random variation in the carbon and cost estimates to assess the impact of error in our estimates on the predicted restoration priorities and outcomes (Supplementary Methods and Results: Appendix S1 and Figure S9).
3. Results
3.1. A Reinforcement Learning Framework for Restoration
We developed an approach to optimize restoration planning building upon the reinforcement learning software CAPTAIN (Conservation Area Prioritization Through Artificial INtelligence) (Silvestro et al. 2022). The implementation is based on a spatially explicit dynamic environment characterized by climatic features, species and their geographic ranges, anthropogenic disturbances, and restoration costs. The disturbance could be any kind of anthropogenic activity including settlements, roads, farming or logging on land and fishing or mining in the sea, while costs represent direct restoration costs and opportunity costs, for example, if a restored area can no longer be used for farming.
The other element of the CAPTAIN framework is an agent (i.e., a virtual representation of a manager or social planner) who monitors the environment and implements a restoration program pursuing one or multiple pre‐defined restoration targets. The action performed by the agent is the selection of areas within which the disturbance is reduced and natural restoration can occur. The agent uses a neural network to map the current state of the environment to a selection of areas for restoration. The environment evolves over time through a number of steps (e.g., years, here set to 30) during which multiple processes take place, including dispersal and mortality among species, changes to climate or disturbances, and the establishment of restoration areas selected by the agent. The intervention of the agent is assigned a reward based on its outcome, for instance its contribution to preventing species extinctions, and models are optimized to maximize the reward based on the outcomes of the agent's actions.
We implemented a multi‐objective reinforcement learning algorithm to evaluate the ability of this approach to optimize restoration toward multiple goals, specifically targeting the net carbon sequestration and reduction in species extinction risks, while incorporating restoration costs (see Section 2, Equations (6), (7), (8), (9), (10), (11), (12)). Here, we measured extinction risks in a 5‐tiered classification inspired by the International Union for Conservation of Nature (IUCN) (IUCN 2023) and ranging from Least Concern to Critically Endangered. The model is highly flexible and can be tuned to optimize other parameters, such as biodiversity metrics based on richness, abundance, and function. We optimized five models to get different spatial prioritizations, which we hereafter refer to as policies (Terryn et al. 2024), using differently weighted rewards. Specifically: Policy (1) focused on carbon, Policy (2) included extinction risk with a small weight, Policy (3) weighted carbon and biodiversity equally, Policy (4) assigned a smaller weight to carbon, and Policy (5) focused on biodiversity (Table 2). All policies shared the same weight on the budget, rewarding the reduction of restoration costs. All policies were constrained to restore the same total area, which makes the biodiversity and carbon outcomes more comparable among policies, but which means that they may differ in implementation costs. We then compared the predicted restoration outcomes of different policies, in terms of selected areas, resulting net carbon sequestration, and impact on species extinction risks. This allowed us to evaluate a gradient of restoration policies ranging from climate‐ to biodiversity‐optimal. In all plots we use a color gradient from brown to green to indicate policies with rewards shifting from carbon to biodiversity.
3.2. Optimization of Restoration Policies and Trade‐Offs
Our simulations of biodiversity used to optimize restoration policies and predict their outcomes revealed that the five policies resulted in substantial spatial differences in the prioritization of restoration areas and led to widely different carbon and biodiversity outcomes. While model training and simulations were based on the same arbitrarily chosen geographic region (Figures S2 and S3), all datasets differed in the distribution of simulated species, threats, and costs in order to cover a wide range of possible real‐world scenarios. Consequently, the placement of the selected restored areas varied among them. However, at the level of each simulation, the policies clearly prioritized different areas (Figure S1).
Policy 1 favored the restoration of areas that resulted in the highest net carbon gain (24% on average after 30 time steps; Figure 1a). In contrast, Policy 5 prioritized areas that contributed the most to reducing species extinction risk, therefore considering threats, species richness, and complementarity of species composition (Figure S1). The result was a more modest carbon gain (just above 11% on average; Figure 1a) but a substantially stronger reduction of the extinction risk compared with Policy 1. For instance, the average number of Critically Endangered species at the end of the simulation was 36% lower under Policy 5 compared to Policy 1 (Figure 1b). Similarly, the number of species classified as Endangered decreased on average by 73% when giving priority to biodiversity over carbon. These experiments also showed that while carbon accumulated through time following near‐linear trajectories, extinction risks followed more complex paths (Figure 1b–f), reflecting the fact that species moved across risk categories and that their conservation status was determined by their ability to re‐colonize restored areas based on their dispersal ability, their sensitivity to disturbance, and connectivity among suitable areas.
FIGURE 1.

Accumulation of carbon (a) and evolution of extinction risks (from Critically Endangered to Least Concern, b, c, d, e, f) across 30 time‐steps averaged over the 100 simulations. The results are shown for Policies 1–5, which differ in the relative weight assigned to carbon sequestration (highest in Policy 1) and reduction of extinction risks (highest in Policy 5). The plots show that when biodiversity is factored in, the carbon accumulation curve reaches lower levels decreasing from 24% to 11% (a). This trend is counterbalanced by a substantial improvement of species conservation status, with a reduction of Critically Endangered and Endangered species of 36% and 73%, respectively, from Policy 1 to 5 (b, c). Meanwhile, the number of Least Concern species increased by 51% (f), highlighting the substantial impact of biodiversity‐targeted restoration in our simulations. The time labels used here are arbitrary.
The distributions of outcomes at the end of the 30 time‐step restoration show a substantial level of consistency across 100 simulations (Figure S6). The net carbon gain in Policy 1, which was trained without considering a biodiversity reward, resulted in an average 2.2‐fold increase (standard deviation among simulations: ±0.26) in carbon storage compared to Policy 5, which did not include carbon in the reward (Figure 2a). Policies with mixed rewards revealed a gradient of outcomes, with net carbon gain gradually decreasing when setting the relative carbon reward to lower values (Figures 1b and 2a). In contrast, the biodiversity outcomes revealed drastic improvements as soon as a non‐zero weight was assigned to the biodiversity reward, with Policies 2–5 all resulting in similar numbers of species with reduced extinction risk (Figure 2b), while Policy 1 resulted in significantly worse impacts on extinction risks compared to most other policies (Figure 2 and Figure S6b). For instance, while on average 30% of the species decreased their extinction risk under Policy 1 (150 out of the 500 simulated species), the percentage rose to 71%–76% (354–378 species) under all other policies (Figure S6b). The fraction of Critically Endangered species remained virtually unchanged under Policy 1, while it decreased on average by ∼29% under Policy 5, and by 27% under all other policies (Figure S6d). When combining all threatened species (CR, EN, and VU), these decreased by about 25% under Policy 1 (i.e., ∼68 species) but dropped by 62%–69% (i.e., 170–190 species) under the others (Figure S6i), thus showing a 2.6‐fold improvement. These results show that even when biodiversity outcomes are given a smaller weight compared to carbon in the optimization of the restoration policy (e.g., Policy 2), the impact on extinction risk reduction is large. Finally, the costs of implementation of the different policies, that is, the cumulative cost of all restored cells, were relatively similar, with changes between 4% and 17% among policies (Figure 2c). The variation results from setting a constraint on the number of restored cells rather than on the total budget, a choice that makes the biodiversity and carbon outcomes more comparable among policies. Yet, the low variation in total costs reflects the fact that the same weight was assigned to the budget reward in the optimization of all policies (see Section 2).
FIGURE 2.

Summary statistics of key outcomes from 100 simulations run under five trained restoration models, Policies 1–5, ranging in their objective from carbon‐only (1), through mixed objectives (2–4), to biodiversity‐only (5). The net carbon (a) represents the net gain in amount of carbon (here expressed in arbitrary units—see Section 2) in the system after 30 time‐steps during which 3000 cells were restored. As the weight assigned to the carbon reward was lowered from 0.5 to 0 in favor of the biodiversity reward, the net carbon gain gradually decreased, although all restoration policies resulted in a net gain. In contrast, the number of species with improved conservation status, that is, lower extinction risk, drastically increased as soon as a non‐zero weight was assigned to the biodiversity reward (Policies 2–5; b). The implementation cost (here expressed in arbitrary units—see Section 2) was similar across all policies (c), reflecting the fact that all models were trained with the same budget reward.
Additional experiments showed that our results are robust to changes in the architecture of the neural network underlying the different policies (Table S3). In contrast, analyses of the simulated data under simple rule‐based selection of the cells to be restored (e.g., prioritizing cells with the highest ratio between number of threatened species and restoration cost) yielded substantially worse outcomes (Table S3; Supplementary Results: Appendix S1). This indicates that the neural network is better able to capture the dynamics of the environment to select restoration areas.
3.3. Carbon Credits and the Impact of Biodiversity Credit Valuations
Our analyses revealed that different restoration policies lead to vastly divergent carbon and biodiversity outcomes while maintaining comparable implementation costs. To evaluate their financial feasibility, we first performed a break‐even analysis assuming a carbon‐only credit market. We set the carbon credit value () such that the cumulative revenue from Policy 1, which maximizes carbon sequestration, exactly matched its total restoration costs (Equation 14; Figure 3a). Under this baseline, all other policies incurred net financial losses, ranging from −10% for Policy 2 to approximately −50% for Policy 5, as they prioritize biodiversity at the expense of carbon sequestration.
FIGURE 3.

Financial outcomes of restoration policies (net percentage loss or gain, i.e., in Equation (15)) under different valuations of biodiversity credits. Negative values indicate a net economic loss where restoration costs exceed the cumulative value of carbon and biodiversity credits, while positive values indicate a net gain. The results are shown when only carbon credits () are included (i.e., ; a), which we used as a baseline of net 0 financial outcome to show how biodiversity credits could contribute value to the simulated restoration programs. The results in (b)–(d) show that when biodiversity credits are valued at and , respectively, all policies become financially viable ().
We then explored the financial impact of bundling carbon and biodiversity credits. Recognizing that there is currently no global consensus on the unit of a biodiversity credit (Sutton et al. 1998; Wauchope et al. 2024), we adopted a performance‐based metric focused on tangible ecological outcomes. For this demonstration, one biodiversity credit () corresponds to one species transitioning from a threatened category to a non‐threatened one by the end of the simulation. Without implying that this should be preferred over alternative definitions, we used this measure to evaluate the potential effects of biodiversity credits on the financial viability of different restoration strategies.
Our experiments with varying relative valuations ( to ) demonstrated that biodiversity credits can effectively level the financial barriers to multi‐objective restoration. When the value of one biodiversity credit was set equal to one carbon credit (), Policy 2—which maintains a carbon‐oriented focus while integrating a moderate biodiversity weight—became profitable, yielding gains above 10%. Further increasing the relative value of biodiversity ( to ) improved the financial outcomes for all policies (Figure 3c,d). Notably, Policy 2 consistently outperformed the others, reflecting the non‐linear biodiversity gains achievable when even moderate weight is given to conservation (Figure S6b).
Finally, we derived the “equivalence value” of a biodiversity credit, that is, the price required for a biodiversity‐aware policy (Policies 2–5) to match the net profit of the carbon‐optimal Policy 1 (Equation 17). We found that Policies 3–5 reached parity with Policy 1 at valuations of , whereas for Policy 2, a value of was sufficient (Table S1). This implies that if a biodiversity credit is valued at just 1.2 times a carbon credit, Policies 1 and 2 become economically indistinguishable.
While we acknowledge these valuations are currently expressed in arbitrary units, our framework provides a reproducible path for deriving the real‐world value of biodiversity credits based on defined ecological targets. In contrast, assigning a null value to biodiversity () leaves limited options to target biodiversity in restoration programs without drastically reducing the financial viability of the policy.
3.4. Restoration Priorities and Trade‐Offs in the Atlantic Forest
We tested our multi‐objective reinforcement learning model on an empirical dataset from the Atlantic Forest—one of the world's most biodiverse and threatened ecosystems (de Lima et al. 2024; Environmental Policy Innovation Center International Advisory Panel on Biodiversity Credits and Forest Trends 2025). Our analysis was based on a recently published dataset (de Lima et al. 2024), which included > 3800 species of trees and shrubs. The data included species distribution models based on approximately 900,000 curated occurrences, an approximation of species‐specific relative contribution to above‐ground carbon, and extinction risks based on the IUCN Red List classification scheme. We also incorporated a proxy for current disturbance levels based on land use, and restoration costs based on accessibility, elevation and value of agricultural production (see Section 2 and Appendix S1; Figure S12).
We ran three models based on Policies 1 (carbon‐focused), 3 (equal weights for carbon and biodiversity), and 5 (biodiversity‐focused), as described above. Simulation experiments carried out outside of the region used for model training showed that the optimized policies can be applied to different contexts with reasonable performance (Figures S1 and S7; Supplementary Methods and Results: Appendix S1). However, they also suggest that models achieve higher and more consistent performance when applied within the region used in their optimization. For this reason, we fine‐tuned the simulation‐trained models with the Atlantic Forest data, to identify the 30% with highest restoration priority (i.e., 4800 of the 16,263 spatial units considered here). The training was based on weighted rewards linked to carbon, species protection, and costs computed as defined in Equations ((6), (7), (8), (9), (10), (11), (12)). As some of the parameters of analysis were not available or incomplete (such as species‐specific growth rates, sensitivity to disturbance, and extinction risks), we performed 30 runs to account for variation in the imputation of missing data (see Section 2).
The selected areas showed substantial differences among policies, highlighting partial mismatches between predicted restoration potential in terms of carbon and biodiversity outcomes (Figure 4a). Specifically, 63% of the areas selected under a carbon‐focused policy overlapped with those selected under a policy that accounted for both carbon and biodiversity, while the overlap dropped to 34.5% when comparing against a policy focused only on biodiversity outcomes. Meanwhile, Policies 3 and 5 overlapped by 53.3%. The estimates obtained across different runs showed that variation in the imputation of missing data has a non‐negligible impact on the results, although selected spatial priorities were generally consistent among replicates. The overlap across runs was 94% for Policy 1, 80% for Policy 3, and 74% for Policy 5, suggesting that missing data had a greater impact on restoration priorities targeting extinction risk compared to carbon (Figure S9).
FIGURE 4.

Results of CAPTAIN optimizations in the Atlantic Forest based on a dataset including habitat suitability maps for 3789 species of trees and shrubs, and proxies for costs and land use. The extent of the considered area is shown in brown, while the cells selected for restoration are shown in green. The data are analyzed under three policies (as defined in Table 2) that were focused on carbon (Policy 1), reduction of extinction risks (Policy 5) or both (Policy 3). All models also used a budget reward for costs and were constrained to restore the same number of grid cells (i.e., 30% of the total). The selected areas, after fine‐tuning the simulation‐trained models with the empirical data, are only partly overlapping, indicating substantial trade‐offs among priorities (a). The predicted outcomes also vary substantially among policies in terms of expected amount of net carbon sequestration (approximated as described in the Section 2) (b) and number of species occurring only outside of protected areas (c). The violin plots show the outcomes of 30 runs that incorporate uncertainties in the imputation of missing data (see Section 2 and Figure S9a–c). All policies selected cells with a wide range of costs (d).
The outcomes of our inference were evaluated through comparisons between the predicted changes in carbon and species extinction risks obtained through restoration and the starting condition (baseline) of current disturbance and threat levels. This comparison is essentially equivalent to a business‐as‐usual counterfactual because the ecosystem is simulated to be at equilibrium with no changes in disturbance, extinction risks, or species abundances through time, in the absence of restoration actions. The results reflected the objectives the models had been optimized for, showing clear trade‐offs between the potential for carbon sequestration through restoration and the ability to protect species and reduce extinction risks. Using species‐specific proxies for above‐ground carbon and calibrating the average carbon per spatial unit to 98.5 tonnes per hectare (Mg C/ha) in the Atlantic Forest (Chave et al. 2005), we found that Policy 1 led to an average predicted net carbon sequestration (approximated as described in the Section 2) ranging from 69.3 to 70 Mg C/ha, across 30 runs (Figure 4b). For comparison, Policies 3 and 5 led to a predicted carbon sequestration of 54.3–58.8 and 30.1–36.9 Mg C/ha, respectively, thus leading to a decrease in the order of 20%–50%.
In contrast, areas selected under Policy 1 left 42–55 species under no protection, a number that dropped to 3–16 under Policy 3 and to 2–8 under Policy 5 (Figure 4c). Under Policy 1, between 15 and 38 of the 420 Critically Endangered and Endangered species included in the dataset (391–453 across the 30 runs) remained in these high‐risk categories at the end of the simulation. This indicates that the selected areas failed to restore a sufficient fraction of species' suitable habitats (set to 25% for CR species and 10% for EN species) required to move to a lower risk category under our thresholds (Table 1). The median number of high‐risk species dropped to 0 under Policy 3, albeit with some degree of variability among runs (0–4), which reduced further under Policy 5, where at most 1 species was left in that group.
The total cost of Policy 3 was about 1% lower than that of Policy 1% and 16% lower than Policy 5. However, all policies showed that the cost of selected cells covers most of the range of affordable cells (i.e., excluding urban areas; Figure 4d). This indicates that the restoration areas optimized by CAPTAIN still included expensive cells despite a negative reward associated with implementation costs.
We carried out sensitivity tests to assess the impact of potential error in carbon estimates and opportunity costs (see Supplementary Methods and Results: Appendix S1). For these analyses we simulated random error in the carbon and cost data and repeated the analyses across 30 runs. The results showed that the selected priority areas remained broadly consistent with those obtained with the original dataset (Figure 4a, Figure S9). However, the simulated error led to a lower average overlap across the priority areas selected in different runs, which decreased to 51%–94% (from the 74% to 94% variation in analyses without error; see Supplementary Methods and Results: Appendix S1). These results show that while some of the selected areas are robust to error in the underlying data, others are more sensitive. This demonstrates the importance of obtaining high‐quality data to reduce uncertainties in real‐world applications of CAPTAIN in restoration planning. The analyses show that our approach can be used to explicitly tackle, rather than ignore, deficiencies and uncertainties inherent in real‐world data, in order to quantify the robustness of the results. Meanwhile, regions where error leads to shifting prioritization highlight areas that might require additional verification of the cost, carbon, and biodiversity data to achieve a robust optimization of restoration planning.
4. Discussion
4.1. Reinforcement Learning as a Tool for Guiding Restoration
Recent work has shown that reinforcement learning holds potential in the context of dynamic ecological settings, for instance in guiding patrolling of protected areas (de Lima et al. 2020), selecting conservation areas (Silvestro et al. 2022; Xu et al. 2022), improving their connectivity (Lapeyrolerie et al. 2022), or modeling ecosystem dynamics (Equihua et al. 2024). Here, we demonstrated the power of multi‐objective reinforcement learning to optimize ecological restoration strategies targeting climate‐ and biodiversity‐positive outcomes while accounting for costs of implementation. Using weighted rewards in the training phase allowed us to calibrate models that explore the boundary of optimality (Lin et al. 2019) between net carbon gain and reduction in extinction risks obtained through restoration, while keeping the extent of restored area constant. This simulation and optimization framework allows us to evaluate the synergies and trade‐offs among restoration policies with different objectives, as also shown in applications in conservation planning (Strannegård et al. 2025; Horn et al. 2026).
As our models are trained within a biologically and environmentally dynamic environment, the solutions incorporate the effects of connectivity among restored areas, which favor migration processes and range expansion underlying the recovery of our simulated species. They can also account for species‐specific features such as climatic niche, potential for carbon sequestration, growth and dispersal rates, and sensitivity to anthropogenic disturbance. The simulations show that the spatial distribution of biodiversity and carbon priorities differs, as observed in real‐world scenarios (Jung et al. 2021; Butikofer et al. 2026). This mismatch means that a restoration program focusing on the maximization of carbon will likely not result in the best biodiversity outcome, and vice versa.
Our analysis of empirical data from the Brazilian Atlantic Forest confirms the trends observed with simulated data. A restoration program optimized exclusively for carbon sequestration fails to adequately protect biodiversity, even when based on natural restoration as opposed to monoculture plantations. We have not considered monocultures in our study, given their well‐known harmful impact on the environment (e.g., (Parr et al. 2024)), and since ecological restoration projects in Brazil do not use single‐species plantations. On the other hand, including the reduction of extinction risks into the optimization will typically also lead to a reduction in net carbon gain. This is because threatened species are not necessarily the fastest‐growing ones (such as the giant timber trees in the genus Cariniana, which take decades to reach maturity) or those with largest biomass (such as the many treelets in the genus Myrcia).
While no solution can thus achieve maximum levels of biodiversity outcomes and carbon sequestration simultaneously, our empirical analysis for the Atlantic Forest shows that the loss of biodiversity protection inflicted by focusing only on carbon would be substantially larger (ca. 7 times more species under no protection) than the reduction of carbon sequestration (ca. 18% decrease) due to the incorporation of biodiversity outcomes (see Policies 1 vs. 3 in Figure 4b,c). The results also indicate that the incorporation of carbon outcomes leads to only a minor loss of biodiversity benefits (Policies 3 vs. 5 in Figure 4b,c). Although our implementation gives substantial flexibility in creating realistic models of the environment's dynamics, ultimately the application of this methodology to real‐world restoration planning depends on the accuracy of the data fed into the model. This will require moving beyond the approximations for cost, land use, and habitat degradation used in our proof‐of‐concept implementation and the integration of additional constraints, for instance linked with property boundaries and restoration feasibility.
Within the limits of available data, our multi‐objective reinforcement learning framework provides a robust tool to explore trade‐offs between the arguably most important objectives of ecological restoration: carbon sequestration and biodiversity uplift, allowing for the optimization of these potentially conflicting metrics. However, we acknowledge that final decisions of where restoration efforts will occur depend on additional considerations other than the ones explored here—such as the local interests and support of Indigenous and local communities, political commitments at national and regional levels, availability of seedling supply, accessibility, human capacity, and other resources.
4.2. The Potential of Bundled Carbon and Biodiversity Credits
Our simulation experiments show that an evaluation of restoration outcomes and their associated costs can be used to forecast the impact of biodiversity credits on program feasibility. We show that the financial sustainability of restoration policies targeting both carbon sequestration and extinction risk is highly sensitive to the valuation of biodiversity relative to carbon. Crucially, our results indicate that even a modest premium assigned to biodiversity outcomes can create incentives for multi‐objective restoration that addresses both carbon sequestration and extinction risks.
In principle, this suggests that high‐integrity carbon credits, which explicitly consider and demonstrate co‐benefits such as biodiversity gain, could help bridge the funding gap for projects that effectively achieve both climate and ecological goals. These carbon credits could be traded at a premium price to compensate for the costs of activities that protect, restore and monitor biodiversity, as well as to generate additional benefits that could be shared within the community involved in the conservation projects. In practice, ecological restoration companies frequently report that such added costs remain too high to be competitive in the global carbon market, pushing restoration projects to marginal, inexpensive land, where ecological potential may be lower. This underscores the urgent need for restoration programs to explicitly consider biodiversity as a central objective (Brancalion et al. 2025), and incorporate funding mechanisms based on robustly measured biodiversity outcomes (Alejo et al. 2025).
Funding based on carbon and biodiversity outcomes can take many forms in practice, including both the voluntary and compliance carbon markets (e.g., EU Emissions Trading System), and the nascent biodiversity credit markets (Antonelli et al. 2024; Gonzalez et al. 2026). The historical financial volatility of carbon markets, even within compliance markets, highlights the crucial role of public policy intervention. Mechanisms such as government‐backed price floors and sovereign guarantees could establish a minimum guaranteed credit value, ensuring stable incentives for multi‐objective restoration even during market fluctuations. At the same time, regulatory measures should follow clearly delimited rules and be independent from political interference to reduce risks of regulatory instability for investors (Croci et al. 2025). Ensuring both price and regulatory stability helps nature credits to become credible instruments to provide carbon and biodiversity benefits.
The framework used in this study does not depend on any particular restoration funding mechanism. In practice, it is likely that a combination of public and private financial tools, including but not limited to nature credits, will be necessary to generate effective results (Boute and Zhang 2019; Zu Ermgassen and Löfqvist 2024).
In our simulations, we define biodiversity credits based on the number of species with decreased extinction risks. We prioritize this species‐based metric over area‐based ones to better capture the “biodiversity uplift” central to emerging credit frameworks: “a measured and evidence‐based unit of positive biodiversity outcome” (Garvey et al. 2025). Although our current analysis focuses on extinction risk, this framework is inherently flexible; it could be adapted to evaluate other metrics such as population recovery, geographic range expansion, or improved ecosystem services (Biodiversity Credit Alliance 2024; Farnsworth et al. 2015). While a simulated setting allows us to quantify restoration at any level of demographic and spatial detail (Silvestro, Goria, Groom, et al. 2025), real‐world implementation will require metrics that can be monitored accurately and cost‐effectively, such as the periodic re‐assessment of extinction risks for indicator species (Antonelli et al. 2023; Durán et al. 2020).
A key distinction between carbon and biodiversity is that sequestered carbon is largely fungible—a ton of CO2 contributes similarly to climate mitigation regardless of where it is stored (barring issues of permanence (Schipper et al. 2020)). In contrast, biodiversity outcomes are inherently context‐dependent (Biodiversity Credit Alliance 2024); the value of recovering a specific species is tied to its rarity, its role in the ecosystem, and its geographic location. Estimating trade‐offs among restoration outcomes and costs can help us define actionable and context‐dependent valuation for in situ biodiversity credits, with their value likely to differ across geographic contexts, spatial extent, and the selection of reference species.
4.3. Proof‐Of‐Concept Application With Real‐World Data
We demonstrated how our framework can be applied to real‐world scenarios by analyzing a large empirical dataset from one of the most highly threatened and diverse ecosystems in the world, considered a Global Biodiversity Hotspot—the Atlantic Forest. Our analysis shows that our reinforcement learning approach can utilize diverse and multi‐dimensional empirical data to identify priority areas for restoration, while balancing potentially conflicting targets and constraints for carbon sequestration, biodiversity outcomes, and implementation costs. While we focused here on the Atlantic Forest, the biodiversity simulation framework and optimization algorithms implemented in CAPTAIN can be adapted to different ecosystems and taxa and across spatial scales, through the parameterization of system‐specific disturbance, recovery, and restoration actions. For instance, the framework has been applied at country or state levels in temperate and boreal forests (Strannegård et al. 2025; Butikofer et al. 2026) and across ecosystems at the continental level (Horn et al. 2026).
Within a 30% restoration scenario in the Atlantic Forest, we identified substantial discrepancies between carbon and biodiversity priorities, pointing to trade‐offs that need to be accounted for to achieve positive change for climate and nature. Policies focused primarily on carbon sequestration tended to prioritize long and cohesive stretches of coastal areas (Figure 4a). This encompassed most of the Brazilian coast within the tropical belt, probably due to the fast growth rates and larger carbon sequestering potential in those areas. In contrast, a biodiversity‐only prioritization (Figure 4a) identified a more scattered set of areas, probably due to the heterogeneous distribution of threatened species across the Atlantic Forest ecosystem.
Implementing spatial optimization tools in practice requires reliable ecological and socioeconomic data at a granularity consistent with the scale of restoration and conservation policies. While our analyses utilize empirical data, they rely on several simplified assumptions and approximations that must be addressed to transition from a proof‐of‐concept to real‐world implementation. In particular, our use of land use as a proxy for anthropogenic disturbance and its modeled effect on plant diversity and abundance could be refined by characterizing disturbance patterns (severity, frequency, and projected trends) within each land use type using high‐resolution data, and modeling their effects on plant diversity and abundance based on principles of plant community dynamics. Similarly, restoration costs—here approximated based on accessibility and land use—could be calibrated by empirical estimates of implementation and opportunity costs (Adams et al. 2019; Groom and Venmans 2023; Alves‐Pinto et al. 2017; Strassburg et al. 2019).
Another challenge in real‐world optimization of restoration strategies is quantifying the degree of uncertainty associated with the underlying data. While robust quantification of uncertainties exists or can be generated for some data (e.g., for climate projections and species distribution models (Brancalion et al. 2021; Knutti and Sedláček 2013)), equivalent estimates are not directly available for other data, such as human footprint and costs. Through our sensitivity tests we demonstrated that such uncertainties can be modeled through simulations, and that, as expected, they do influence the inferred restoration priorities (Figure S9). As the amount and quality of spatial data continue to improve through data compilation, remote sensing, and emerging Earth foundation models (Buisson et al. 2010), it is likely that the availability of robust estimates of anthropogenic disturbance, costs, and other socio‐economic variables will improve over time and allow increasingly detailed restoration planning. Potentially, the quantified uncertainties in restoration priorities can also be used to guide the acquisition of additional data targeting areas where new data are most likely to affect the prioritization outcomes.
The simulated restoration process through natural recolonization based on dispersal and growth could also be improved based on the findings of ongoing restoration experiments to inform how different species can be re‐introduced in restored habitats to accelerate the recovery of diversity and ecosystem functions (Audino et al. 2017; Feng et al. 2026; Wang et al. 2022; Selwyn et al. 2023).
Crucially, the objectives, needs, and knowledge of Indigenous and local communities, which are essential for fair and effective conservation of nature (Erbaugh et al. 2020, Simões et al. 2025), should be integrated into the decision‐making process. This could be achieved by restricting the range of candidate sites available for restoration to only comprise those that may indeed be available for that purpose, by imposing particular constraints on the optimization objectives deriving from the custodians of land, or by only including in the analyses species that are available and adequate for use in restoration. Work towards that goal is already underway in CAPTAIN (Silvestro, Goria, Rau, et al. 2025).
5. Conclusion
The controlled simulations and empirical analyses performed here demonstrate the power and practical utility of a reinforcement learning framework to select priority areas for ecological restoration that substantially benefit climate and biodiversity. Focusing on a single restoration objective would result in missed opportunities (Brancalion et al. 2025), whereas integrating both carbon capture and biodiversity conservation in a multi‐objective prioritization allows us to predict the outcomes of different restoration policies.
Our model offers a way to explore how costs may be distributed between key financial mechanisms such as carbon credits and biodiversity credits (Antonelli et al. 2024), which in turn can serve as a tool to estimate how a biodiversity crediting system could help achieve both climate and biodiversity targets. Future model developments could include additional funding sources, such as payments for ecosystem services (clean water and pollination) that increase over time as well‐functioning ecosystems are recovered.
Ecosystem restoration is increasingly recognized as a crucial opportunity in the conservation of nature and stabilization of Earth's climate (Seddon et al. 2020; Strassburg et al. 2020). The use of robust models and AI‐based optimization, informed by the increasing availability of remote‐sensed data and ground surveys, can help us deal with the complexity of biodiversity and its spatial and temporal dynamics, thus playing a critical role in guiding society's efforts to tackle the twin crises of biodiversity loss and climate change.
Author Contributions
Daniele Silvestro: conceptualization, investigation, writing – original draft, methodology, validation, visualization, writing – review and editing, software, formal analysis. Stefano Goria: conceptualization, writing – original draft, methodology, software. E‐Ping Rau: conceptualization, methodology, formal analysis, writing – review and editing, writing – original draft, data curation. Renato Augusto Ferreira de Lima: data curation, writing – review and editing. Ben Groom: conceptualization, writing – review and editing. Piotr Jacobsson: methodology, formal analysis. Thomas Sterner: conceptualization, writing – review and editing. Alexandre Antonelli: conceptualization, investigation, writing – review and editing, project administration, funding acquisition, writing – original draft.
Funding
A.A., B.G., T.S., and D.S. are members of the BIOPATH research programme funded by the Swedish Foundation for Strategic Environmental Research MISTRA (F 2022/1448). A.A. and E‐P.R. are funded by the Swedish Research Council (VR: 2019‐05191, 2024‐04303) and the Royal Botanic Gardens, Kew; B.G. is funded by Dragon Capital and the NERC/UKRI projects BIOADD (ref: NE/X002292/1), BIOESG (NE/X016560/1) and RENEW (ref: NE/W004941/1); D.S. is funded by ETH Zurich and the PEACE research program funded by Swiss National Science Foundation and the European Biodiversity Partnership (www.biodiversa.eu/2026/04/03/peace; 31BD‐0_233170); T.S. is funded by the Kamprad Foundation. R.A.F.L. received funding from the São Paulo Research Foundation (FAPESP) grant #13/08722‐5 and by the European Union's Horizon 2020 research and innovation program under the Marie Sklodowska‐Curie grant agreement no. 795114.
Conflicts of Interest
D.S., A.A., T.S., and S.G. are co‐founders of Captain Technologies LTD.
Supporting information
Table S1: Biodiversity credit valuation estimated to reach the same net financial gain as in Policy 1 (focused on carbon sequestration) across all other policies (with increasing weight associated with extinction risks), computed based on Equation (17). The values are calculated across 100 simulations based on the reference region (on which models were trained).
Table S2: Biodiversity credit valuation estimated to reach the same net financial gain as in Policy 1 across all policies, computed based on Equation (17). The values are calculated across 100 simulations based on a set of random regions differing in size and location from the region used to train the policies (Figure S3).
Table S3: Summary of the outcomes of 100 simulations performed under different restoration models. Nonparametric models 1–3 (see Supplementary Methods and Results: Appendix S1) were based on simple metrics targeting carbon and costs (1); carbon, biodiversity, and costs (2); and biodiversity and costs (3). The other models (Policies 1–5) were based on CAPTAIN‐trained neural networks of 3 and 2 nodes in two hidden layers (small NN) and 16 and 8 nodes (larger NN). The effectively protected species quantify the number of species that started from a threatened category (CR, EN, and VU) at the beginning of the simulations but moved to a non‐threatened category (NT and LC) after restoration, while the species with improved status are species that moved to a lower extinction risk. All values are given as median among simulations and 2.5% and 97.5% quantiles rounded to integer.
Figure S1: Changes in spatial restoration priorities depending on the relative weights of carbon (highest in Policy 1) and biodiversity (highest in Policy 5). (a) Example of the restored areas in one of the 100 simulations performed on the reference region used to optimize the models (Figure S2) selected under five policies (Table 2). Note that the placement of the restored areas depends on the specific conditions of the simulated dataset. Restored areas will therefore vary among simulations, reflecting the randomly initialized distribution of species, disturbance, and costs. (b) Inferred restoration priorities obtained from simulations based on another, randomly selected region, showing similar discrepancies in spatial prioritization depending on the relative importance attributed to carbon versus biodiversity outcomes (Figure S3).
Figure S2: Map of the arbitrary area used for model training and subsequent simulations with examples of the empirical and simulated data used in our CAPTAIN runs. The physical properties of the area are based on the real topography and climate of the selected portion of Central and South America (a). The area was gridded into 128 × 128 cells and included empirical climate and elevation layers, while the “natural” state of biodiversity and carbon (b) and the disturbance patterns and costs (c) were simulated and changed randomly across simulations. Species richness and carbon maps are obtained after stacking 500 simulated species (see examples in d). Note that only the map in panel (a) is based on real data, while all other panels show simulated data. Similar simulations were repeated on 100 additional random maps of different size and location (Figure S3).
Figure S3: Maps showing the 100 randomly sampled rasters used as the basis for our simulations. The raster highlighted in red indicates the region used to train the models and to run the simulations based on a fixed geographic extent (see Figure S2 and Section 2).
Figure S4: Training rewards across epochs based on a small policy network (see Section 2).
Figure S5: Training rewards across epochs based on a larger policy network (see Section 2).
Figure S6: Summary statistics of key outcomes from 100 simulations run under five trained restoration models, Policies 1–5. The net carbon (a) represents the net gain in amount of carbon (here expressed in arbitrary units—see Section 2) in the system after 30 time‐steps during which 3000 cells were restored. As the weight assigned to the carbon reward was lowered from 0.5 to 0 in favor of the biodiversity reward, the net carbon gain gradually decreased, although all restoration policies resulted in a net gain. In contrast, the number of species with improved conservation status, that is, lower extinction risk, drastically increased as soon as a non‐zero weight was assigned to the biodiversity reward (Policies 2–5; b). The implementation cost was similar across all policies (c), reflecting the fact that all models were trained with the same weight on the budget reward. Plots (d–h) show the difference between the number of species in each extinction risk category, going from highest risk (CR) to lowest (LC), at the beginning of the simulation and the corresponding number at the end of the simulation. Thus, negative numbers, here observed in (d)–(g) mean that the number of species in the corresponding risk categories decreased. Positive numbers, here mostly observed in (h), mean that the number of species in these categories increased. These plots show that all policies that considered biodiversity rewards in the optimization (Policies 2–5) resulted in substantial reductions of threatened species compared to a policy targeting carbon only. This is also evident when lumping together all threatened species (i).
Figure S7: Summary of 100 simulations from random locations and variable resolution (Figure S3) run under five trained restoration models, Policies 1–5. The general trends in terms of net carbon gain (decreasing from Policy 1 to 5; a), number of species with improved conservation status (higher from Policies 2–5; b), and costs (roughly constant across all policies) are similar to those observed across the simulations based on a fixed geographic extent, upon which the models were optimized. However the patterns are less consistent when looking at changes in species number across risk category (from CR to LC; d–h) or moving from a threatened to a non‐threatened classification (i). These results suggest that the models (here trained on a single region) could be re‐optimized or fine‐tuned to best fit different regions.
Figure S8: Financial outcomes of restoration policies (net percentage loss or gain, that is, in Equation (15)) under different valuations of biodiversity credits, obtained from running the simulations across 100 random regions with different spatial location, extent and resolution. Negative values indicate a net economic loss where restoration costs exceed the cumulative value of carbon and biodiversity credits, while positive values indicate a net gain. The results are shown when only carbon credits () are included (i.e., ; a), which we used as a baseline of net 0 financial outcome to show how biodiversity credits could contribute value to the simulated restoration programs. The results in (b)–(d) show that when biodiversity credits are valued at , all policies become financially viable ().
Figure S9: Restoration priorities in the Atlantic Forest: results for the three policies show the frequency at which each grid cell was selected for restoration (green‐blue shades) across 30 replicates, where growth rates, species sensitivity to disturbance, and missing extinction risk assessments were randomly imputed (a–c). Sensitivity tests show the impact of simulated error in per‐species carbon (d–f) and in the spatial distribution of restoration costs (g–i).
Figure S10: Correlation coefficients for the seven climatic variables (Fick and Hijmans 2017) retained for species distribution modeling.
Figure S11: Violin plots and box plots of the five performance metrics of species distribution models across (a) all species and (b) retained species after applying exclusion criteria (sensitivity < 0.7, specificity < 0.7, True Skill Statistic < 0.5, and Area Under Curve < 0.8).
Figure S12: Approximate values of (a) restoration costs, (b) anthropogenic disturbance, (c) simulated total carbon in the absence of disturbance, and (d) simulated species richness based on stacking the species distribution models. The values are approximated as described in the Section 2.
Figure S13: Example of Perlin noise applied as random cost multiplier to evaluate the impact of error in the cost layer on the inferred restoration priorities. The multiplier is rescaled to alter the cost layer by ±25%.
Acknowledgments
Generative AI statement: Generative AI was not used in producing the contents of this article. Open access publishing facilitated by Eidgenossische Technische Hochschule Zurich, as part of the Wiley ‐ Eidgenossische Technische Hochschule Zurich agreement via the Consortium Of Swiss Academic Libraries.
Contributor Information
Daniele Silvestro, Email: daniele.silvestro@bsse.ethz.ch.
Alexandre Antonelli, Email: a.antonelli@kew.org.
Data Availability Statement
All the code and data used to reproduce the results shown in this study are available on Dryad: https://doi.org/10.5061/dryad.r7sqv9svg. The latest version of the CAPTAIN software is available here: https://www.captain‐project.net.
References
- Adams, V. M. , Iacona G. D., and Possingham H. P.. 2019. “Weighing the Benefits of Expanding Protected Areas Versus Managing Existing Ones.” Nature Sustainability 2, no. 5: 404–411. 10.1038/s41893-019-0275-5. [DOI] [Google Scholar]
- Alejo, C. , Luers A., Ventimiglia A., and Matthews H. D.. 2025. “Maximizing Nature‐Based Solutions Using Artificial Intelligence to Align Global Biodiversity, Climate, and Water Targets. bioRxiv.” Pages 2025–01.
- Alves‐Pinto, H. N. , Latawiec A. E., Strassburg B. B. N., et al. 2017. “Reconciling Rural Development and Ecological Restoration: Strategies and Policy Recommendations for the Brazilian Atlantic Forest.” Land Use Policy 60: 419–426. [Google Scholar]
- Antonelli, A. , Dhanjal‐Adams K. L., and Silvestro D.. 2023. “Integrating Machine Learning, Remote Sensing and Citizen Science to Create an Early Warning System for Biodiversity.” Plants, People, Planet 5, no. 3: 307–316. 10.1002/ppp3.10337. [DOI] [Google Scholar]
- Antonelli, A. , Rueda X., Calcagno R., and Kalunda P. N.. 2024. “How Biodiversity Credits Could Help to Conserve and Restore Nature.” Nature 634, no. 8036: 1045–1049. 10.1038/d41586-024-03475-2. [DOI] [PubMed] [Google Scholar]
- Apicella, A. , Donnarumma F., Isgrò F., and Prevete R.. 2021. “A Survey on Modern Trainable Activation Functions.” Neural Networks 138: 14–32. 10.1016/j.neunet.2021.01.026. [DOI] [PubMed] [Google Scholar]
- Audino, L. D. , Murphy S. J., Zambaldi L., Louzada J., and Comita L. S.. 2017. “Drivers of Community Assembly in Tropical Forest Restoration Sites: Role of Local Environment, Landscape, and Space.” Ecological Applications 27, no. 6: 1731–1745. 10.1002/eap.1562. [DOI] [PubMed] [Google Scholar]
- Bache, S. M. , and Wickham H.. 2022. “magrittr: A Forward‐Pipe Operator for R.”
- Bengtsson, H. 2021. “A Unifying Framework for Parallel and Distributed Processing in R Using Futures.” R Journal 13, no. 2: 208–227. 10.32614/RJ-2021-048. [DOI] [Google Scholar]
- Biodiversity Credit Alliance . 2024. “Definition of a Biodiversity Credit. Issue Paper.” www.biodiversitycreditalliance.org/wp‐content/uploads/2024/05/Definition‐of‐a‐Biodiversity‐Credit‐Rev‐220524.pdf.
- Bordin, K. M. , Esquivel‐Muelbert A., Klipel J., et al. 2023. “No Relationship Between Biodiversity and Forest Carbon Sink Across the Subtropical Brazilian Atlantic Forest.” Perspectives in Ecology and Conservation 21, no. 2: 112–120. 10.1016/j.pecon.2023.02.003. [DOI] [Google Scholar]
- Boute, A. , and Zhang H.. 2019. “Fixing the Emissions Trading Scheme: Carbon Price Stability in the Eu and China.” European Law Journal 25, no. 3: 333–347. 10.1111/eulj.12307. [DOI] [Google Scholar]
- Brancalion, P. H. S. , Guillemot J., César R. G., et al. 2021. “The Cost of Restoring Carbon Stocks in Brazil's Atlantic Forest.” Land Degradation & Development 32, no. 2: 830–841. 10.1002/ldr.3764. [DOI] [Google Scholar]
- Brancalion, P. H. S. , Hua F., Joyce F. H., Antonelli A., and Holl K. D.. 2025. “Moving Biodiversity From an Afterthought to a Key Outcome of Forest Restoration.” Nature Reviews Biodiversity 1: 1–14. [Google Scholar]
- Brancalion, P. H. S. , Meli P., Tymus J. R. C., et al. 2019. “What Makes Ecosystem Restoration Expensive? A Systematic Cost Assessment of Projects in Brazil.” Biological Conservation 240: 108274. 10.1016/j.biocon.2019.108274. [DOI] [Google Scholar]
- Breiner, F. T. , Guisan A., Bergamini A., and Nobis M. P.. 2015. “Overcoming Limitations of Modelling Rare Species by Using Ensembles of Small Models.” Methods in Ecology and Evolution 6, no. 10: 1210–1218. 10.1111/2041-210X.12403. [DOI] [Google Scholar]
- Buisson, L. , Thuiller W., Casajus N., Lek S., and Grenouillet G.. 2010. “Uncertainty in Ensemble Forecasting of Species Distribution.” Global Change Biology 16, no. 4: 1145–1157. 10.1111/j.1365-2486.2009.02000.x. [DOI] [Google Scholar]
- Burgess, N. D. , Ali N., Bedford J., et al. 2024. “Global Metrics for Terrestrial Biodiversity.” Annual Review of Environment and Resources 49, no. 1: 673–709. 10.1146/annurev-environ-121522-045106. [DOI] [Google Scholar]
- Butikofer, L. , Silvestro D., Rubio Teso L., et al. 2026. “Future‐Proofing Agrobiodiversity: Climate and Niche‐Aware Conservation Planning Using Reinforcement Learning. bioRxiv.” 10.64898/2026.05.04.722509. [DOI]
- Chave, J. , Andalo C., Brown S., et al. 2005. “Tree Allometry and Improved Estimation of Carbon Stocks and Balance in Tropical Forests.” Oecologia 145, no. 1: 87–99. 10.1007/s00442-005-0100-x. [DOI] [PubMed] [Google Scholar]
- Croci, E. , Lucchitta B., and Cusa M.. 2025. “Biodiversity Credits Schemes: A Comparative Analysis.” Journal of Cleaner Production 523: 146382. 10.1016/j.jclepro.2025.146382. [DOI] [Google Scholar]
- Cysneiros, V. C. , Pelissari A. L., Gaui T. D., et al. 2020. “Modeling of Tree Height–Diameter Relationships in the Atlantic Forest: Effect of Forest Type on Tree Allometry.” Canadian Journal of Forest Research 50, no. 12: 1289–1298. 10.1139/cjfr-2020-0060. [DOI] [Google Scholar]
- de Lima, R. A. F. , Dauby G., de Gasper A. L., et al. 2024. “Comprehensive Conservation Assessments Reveal High Extinction Risks Across Atlantic Forest Trees.” Science 383, no. 6679: 219–225. [DOI] [PubMed] [Google Scholar]
- de Lima, R. A. F. , Oliveira A. A., Pitta G. R., et al. 2020. “The Erosion of Biodiversity and Biomass in the Atlantic Forest Biodiversity Hotspot.” Nature Communications 11, no. 1: 6347. 10.1038/s41467-020-20217-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Deutz, A. , Heal G. M., Niu R., et al. 2020. “Financing Nature: Closing the Global Biodiversity Financing Gap.” Paulson Institute, the Nature Conservancy, and the Cornell Atkinson Center for Sustainability 256: 1–262. [Google Scholar]
- Di Sacco, A. , Hardwick K. A., Blakesley D., et al. 2021. “Ten Golden Rules for Reforestation to Optimize Carbon Sequestration, Biodiversity Recovery and Livelihood Benefits.” Global Change Biology 27, no. 7: 1328–1348. [DOI] [PubMed] [Google Scholar]
- Drupp, M. A. , Hänsel M. C., Fenichel E. P., et al. 2024. “Accounting for the Increasing Benefits From Scarce Ecosystems.” Science 383, no. 6687: 1062–1064. [DOI] [PubMed] [Google Scholar]
- Durán, A. P. , Green J. M. H., West C. D., et al. 2020. “A Practical Approach to Measuring the Biodiversity Impacts of Land Conversion.” Methods in Ecology and Evolution 11, no. 8: 910–921. 10.1111/2041-210X.13427. [DOI] [Google Scholar]
- Environmental Policy Innovation Center International Advisory Panel on Biodiversity Credits and Forest Trends . 2025. Learnings From Government‐Led Approaches to Nature Credit Markets. IAPB. [Google Scholar]
- Equihua, J. , Beckmann M., and Seppelt R.. 2024. “Connectivity Conservation Planning Through Deep Reinforcement Learning.” Methods in Ecology and Evolution 15, no. 4: 779–790. 10.1111/2041-210X.14300. [DOI] [Google Scholar]
- Erbaugh, J. T. , Pradhan N., Adams J., et al. 2020. “Global Forest Restoration and the Importance of Prioritizing Local Communities.” Nature Ecology & Evolution 4, no. 11: 1472–1476. 10.1030/s41559-020-01282-2. [DOI] [PubMed] [Google Scholar]
- Estopinan, J. , Bonnet P., Servajean M., Munoz F., and Joly A.. 2024. “Modelling Species Distributions With Deep Learning to Predict Plant Extinction Risk and Assess Climate Change Impacts.” arXiv preprint arXiv:2401.05470.
- Eyres, A. , Ball T., Dales M., et al. 2024. “Life: A Metric for Quantitively Mapping the Impact of Land‐Cover Change on Global Extinctions.” Earth and Environmental Science. 10.33774/coe-2023-gpn4p-v3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Farnsworth, K. D. , Adenuga A. H., and De Groot R. S.. 2015. “The Complexity of Biodiversity: A Biological Perspective on Economic Valuation.” Ecological Economics 120: 350–354. 10.1016/j.ecolecon.2015.10.003. [DOI] [Google Scholar]
- Farr, T. G. , and Kobrick M.. 2000. “Shuttle Radar Topography Mission (SRTM).”
- Feng, Z. , Atzberger C., Jaffer S., et al. 2026. “Tessera: Temporal Embeddings of Surface Spectra for Earth Representation and Analysis. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.” 34818–34831.
- Fick, S. E. , and Hijmans R. J.. 2017. “WorldClim 2: New 1‐Km Spatial Resolution Climate Surfaces for Global Land Areas: New Climate Surfaces for Global Land Areas.” International Journal of Climatology 37, no. 12: 4302–4315. [Google Scholar]
- Garvey, J. , Grigoriadis V., Flannery D., et al. 2025. “Designing Financial Instruments for Land‐Based Ecological Restoration: A Review and Future Research Agenda.” Cleaner Production Letters 8: 100089. 10.1016/j.clpl.2024.100089. [DOI] [Google Scholar]
- Gombin, J. , Vaidyanathan R., and Agafonkin V.. 2020. “Concaveman: A Very Fast 2D Concave Hull Algorithm.”
- Gonzalez, A. , August T., Bailey S., et al. 2026. “From Data to Decisions: Toward a Biodiversity Monitoring Standards Framework.” Proceedings of the National Academy of Sciences of the United States of America 123, no. 10: e2519347123. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Groom, B. , and Venmans F.. 2023. “The Social Value of Offsets.” Nature 619, no. 7971: 768–773. 10.1038/s41586-023-06153-x. [DOI] [PubMed] [Google Scholar]
- Gumbs, R. , Gray C. L., Böhm M., et al. 2023. “The edge2 Protocol: Advancing the Prioritisation of Evolutionarily Distinct and Globally Endangered Species for Practical Conservation Action.” PLoS Biology 21, no. 2: e3001991. 10.1371/journal.pbio.3001991. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Hernangómez, D. 2023. “Using the Tidyverse With Terra Objects: The Tidyterra Package.” Journal of Open Source Software 8, no. 91: 5751. 10.21105/joss.05751. [DOI] [Google Scholar]
- Hijmans, R. J. 2025. “terra: Spatial Data Analysis.”
- Hijmans, R. J. , Barbosa M., Ghosh A., and Mandel A.. 2024. “geodata: Download Geographic Data.”
- Holl, K. D. , and Brancalion P. H. S.. 2020. “Tree Planting Is Not a Simple Solution.” Science 368, no. 6491: 580–581. 10.1126/science.aba8232. [DOI] [PubMed] [Google Scholar]
- Horn, K. H. , Silvestro D., Wallis C. I. B., et al. 2026. “A Data‐Driven Framework for Forest Conservation Prioritisation: From Citizen Science Data to Protected Area Designation.” Ecological Informatics 96: 103794. [Google Scholar]
- Hua, F. , Bruijnzeel L. A., Meli P., et al. 2022. “The Biodiversity and Ecosystem Service Contributions and Trade‐Offs of Forest Restoration Approaches.” Science 376, no. 6595: 839–844. 10.1126/science.abl4649. [DOI] [PubMed] [Google Scholar]
- IUCN . 2023. “Red List of Threatened Species. Red List of the International Union for the Conservation of Nature.” http://www.iucnredlist.org.
- IUCN Standards and Petitions Committee . 2025. “Guidelines for Using the IUCN Red List Categories and Criteria. Version 16.” http://www.iucnredlist.org/documents/RedListGuidelines.pdf.
- Jung, M. , Arnell A., De Lamo X., et al. 2021. “Areas of Global Importance for Conserving Terrestrial Biodiversity, Carbon and Water.” Nature Ecology & Evolution 5, no. 11: 1499–1509. [DOI] [PubMed] [Google Scholar]
- Karger, D. N. , Nobis M. P., Normand S., Graham C. H., and Zimmermann N. E.. 2023. “Chelsa‐trace21k–High‐Resolution (1 Km) Downscaled Transient Temperature and Precipitation Data Since the Last Glacial Maximum.” Climate of the Past 19, no. 2: 439–456. 10.5194/cp-19-439-2023. [DOI] [Google Scholar]
- Kattge, J. , Diaz S., Lavorel S., et al. 2011. “Try—A Global Database of Plant Traits.” Global Change Biology 17, no. 9: 2905–2935. 10.1111/j.1365-2486.2011.02451.x. [DOI] [Google Scholar]
- Knutti, R. , and Sedláček J.. 2013. “Robustness and Uncertainties in the New cmip5 Climate Model Projections.” Nature Climate Change 3, no. 4: 369–373. 10.1038/nclimate1716. [DOI] [Google Scholar]
- Lamahewage, S. H. G. , Witharana C., Riemann R., Fahey R., and Worthley T.. 2025. “Aboveground Biomass Estimation Using Multimodal Remote Sensing Observations and Machine Learning in Mixed Temperate Forest.” Scientific Reports 15, no. 1: 31120. 10.1038/s41598-025-15585-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Lang, N. , Jetz W., Schindler K., and Wegner J. D.. 2023. “A High‐Resolution Canopy Height Model of the Earth.” Nature Ecology & Evolution 7, no. 11: 1778–1789. 10.1038/s41559-023-02206-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Lapeyrolerie, M. , Chapman M. S., Norman K. E. A., and Boettiger C.. 2022. “Deep Reinforcement Learning for Conservation Decisions.” Methods in Ecology and Evolution 13, no. 11: 2649–2662. 10.1111/2041-210X.13954. [DOI] [Google Scholar]
- Lin, X. , Zhen H.‐L., Li Z., Zhang Q.‐F., and Kwong S.. 2019. “Pareto Multi‐Task Learning. Advances in Neural Information Processing Systems.” 32.
- Lopez‐Gonzalez, G. , Lewis S. L., Burkitt M., and Phillips O. L.. 2011. “Forestplots. Net: A Web Application and Research Tool to Manage and Analyse Tropical Forest Plot Data.”
- Mair, L. , Bennun L. A., Brooks T. M., et al. 2021. “A Metric for Spatially Explicit Contributions to Science‐Based Species Targets.” Nature Ecology & Evolution 5, no. 6: 836–844. 10.1038/s41559-021-01432-0. [DOI] [PubMed] [Google Scholar]
- MapBiomas . 2024a. “MapBiomas Collection 10: Land Cover and Use—Brazil Coverage.”
- MapBiomas . 2024b. “MapBiomas Brasil Collection 9.” https://brasil.mapbiomas.org/en/map/colecao‐9/.
- MapBiomas . 2025. https://storage.googleapis.com/mapbiomas‐public/initiatives/brasil/collection_10/lulc/coverage/brazil_coverage_2024.tif.
- Maron, M. , von Hase A., Quétier F., Sonter L. J., Theis S., and zu Ermgassen S. O. S. E.. 2025. “Biodiversity Offsets, Their Effectiveness and Their Role in a Nature Positive Future.” Nature Reviews Biodiversity 1: 1–14. [Google Scholar]
- Nelson, A. , Weiss D. J., van Etten J., Cattaneo A., McMenomy T. S., and Koo J.. 2019. “A Suite of Global Accessibility Indicators.” Scientific Data 6, no. 1: 266. 10.1038/s41597-019-0265-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Olson, D. M. , and Dinerstein E.. 2002. “The Global 200: Priority Ecoregions for Global Conservation.” Annals of the Missouri Botanical Garden 89, no. 2: 199. 10.2307/3298564. [DOI] [Google Scholar]
- Parr, C. L. , Te Beest M., and Stevens N.. 2024. “Conflation of Reforestation With Restoration Is Widespread.” Science 383, no. 6684: 698–701. 10.1126/science.adj0899. [DOI] [PubMed] [Google Scholar]
- Pebesma, E. 2018. “Simple Features for R: Standardized Support for Spatial Vector Data.”
- Pebesma, E. , and Bivand R.. 2023. Spatial Data Science: With Applications in R. Chapman and Hall/CRC. 10.1201/9780429459016. [DOI] [Google Scholar]
- Pimiento, C. , Leprieur F., Silvestro D., et al. 2020. “Functional Diversity of Marine Megafauna in the Anthropocene.” Science Advances 6: eaay7650. 10.1126/sciadv.aay7650. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Possingham, H. P. , Bode M., and Klein C. J.. 2015. “Optimal Conservation Outcomes Require Both Restoration and Protection.” PLoS Biology 13, no. 1: e1002052. 10.1371/journal.pbio.1002052. [DOI] [PMC free article] [PubMed] [Google Scholar]
- R Core Team . 2024. “R: A Language and Environment for Statistical Computing.”
- Salimans, T. , Ho J., Chen X., Sidor S., and Sutskever I.. 2017. “Evolution Strategies as a Scalable Alternative to Reinforcement Learning.” arXiv Preprint arXiv:1703.03864.
- Schipper, A. M. , Hilbers J. P., Meijer J. R., et al. 2020. “Projecting Terrestrial Biodiversity Intactness With Globio 4.” Global Change Biology 26, no. 2: 760–771. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Seddon, N. , Chausson A., Berry P., Girardin C. A. J., Smith A., and Turner B.. 2020. “Understanding the Value and Limits of Nature‐Based Solutions to Climate Change and Other Global Challenges.” Philosophical Transactions of the Royal Society, B: Biological Sciences 375, no. 1794: 20190120. 10.1098/rstb.2019.0120. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Selwyn, M. , Pino J., and Espelta J. M.. 2023. “Disentangling the Importance of Intrinsic and Extrinsic Seed Dispersal Factors in Forest Restoration Success: A Global Review.” Restoration Ecology 31, no. 4: e13868. 10.1111/rec.13868. [DOI] [Google Scholar]
- Silvestro, D. , Goria S., Groom B., Sterner T., and Antonelli A.. 2025. “The 30 by 30 Biodiversity Commitment and Financial Disclosure: Metrics Matter.” Current Opinion in Environmental Sustainability 77: 101587. 10.1016/j.cosust.2025.101587. [DOI] [Google Scholar]
- Silvestro, D. , Goria S., Rau E‐P., Sterner T., and Antonelli A.. 2025. “Reconciling Top‐Down and Bottom‐Up Conservation Priorities With Reinforcement Learning From Human Feedback.” Environmental Research Letters. 10.1088/1748-9326/aea2ab. [DOI] [Google Scholar]
- Silvestro, D. , Goria S., Sterner T., and Antonelli A.. 2022. “Improving Biodiversity Protection Through Artificial Intelligence.” Nature Sustainability 5, no. 5: 415–424. 10.1038/s41893-022-00851-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Simões, L. H. P. , Peña‐Claros M., Guillemot J., et al. 2025. “Recovery of Tree Species Functional Composition in Eucalypt Plantations With Natural Regeneration Differs Among Canopy Strata.” Forest Ecology and Management 594: 122952. 10.1016/j.foreco.2025.122952. [DOI] [Google Scholar]
- Strannegård, C. , Palak M., Engsner N., Stocco A., Antonelli A., and Silvestro D.. 2025. “Predicting Ecosystem Resilience Using Multi‐Agent Reinforcement Learning. bioRxiv.” Pages 2025–06.
- Strassburg, B. B. , Beyer H. L., Crouzeilles R., et al. 2019. “Strategic Approaches to Restoring Ecosystems Can Triple Conservation Gains and Halve Costs.” Nature Ecology & Evolution 3, no. 1: 62–70. [DOI] [PubMed] [Google Scholar]
- Strassburg, B. B. , Iribarrem A., Beyer H. L., et al. 2020. “Global Priority Areas for Ecosystem Restoration.” Nature 586, no. 7831: 724–729. [DOI] [PubMed] [Google Scholar]
- Sutton, R. S. , Barto A. G., and Barto A.. 1998. Reinforcement Learning: An Introduction, Volume 1. MIT press. [Google Scholar]
- Sutton, R. S. , McAllester D., Singh S., and Mansour Y.. 1999. “Policy Gradient Methods for Reinforcement Learning With Function Approximation.” Advances in Neural Information Processing Systems 12: 1057–1063. [Google Scholar]
- Terryn, L. , Calders K., Meunier F., et al. 2024. “New Tree Height Allometries Derived From Terrestrial Laser Scanning Reveal Substantial Discrepancies With Forest Inventory Methods in Tropical Rainforests.” Global Change Biology 30, no. 8: e17473. [DOI] [PubMed] [Google Scholar]
- Torabi, N. , and Bekessy S. A.. 2015. “Bundling and Stacking in Bio‐Sequestration Schemes: Opportunities and Risks Identified by Australian Stakeholders.” Ecosystem Services 15: 84–92. 10.1016/j.ecoser.2015.08.001. [DOI] [Google Scholar]
- U.S. Geological Survey . 1996. “GTOPO30: Global 30‐Arc‐Second Elevation Data.”
- Valavi, R. , Guillera‐Arroita G., Lahoz‐Monfort J. J., and Elith J.. 2021. “Predictive Performance of Presence‐Only Species Distribution Models: A Benchmark Study With Reproducible Code.” Ecological Monographs 92, no. 1: e01486. 10.1002/ecm.1486. [DOI] [Google Scholar]
- Velazco, S. J. E. , Rose B., Andrade A. F. A., Minoli I., and Franklin J.. 2022. “Flexsdm: An R Package for Supporting a Comprehensive and Flexible Species Distribution Modelling Workflow.”
- Wang, T. , Dong L., and Liu Z.. 2022. “Factors Driving Native Tree Species Restoration in Plantations and Tree Structure Conversion in Chinese Temperate Forests.” Forest Ecology and Management 507: 119989. 10.1016/j.foreco.2021.119989. [DOI] [Google Scholar]
- Warren, D. , and Dinnage R.. 2025. “ENMTools: Analysis of Niche Evolution Using Niche and Distribution Models.”
- Wauchope, H. S. , Zu Ermgassen S. O. S. E., Jones J. P. G., Carter H., Schulte to Bühne H., and Milner‐Gulland E. J.. 2024. “What Is a Unit of Nature? Measurement Challenges in the Emerging Biodiversity Credit Market.” Proceedings. Biological sciences 291, no. 2036: 20242353. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Weltbiodiversitätsrat IPBES . 2019. Intergovernmental Science‐Policy Platform on Biodiversity and Ecosystem Services. Summary for Policy Makers of the Global Assessment Report on Biodiversity and Ecosystem Services of the Intergovernmental Science‐Policy Platform on Biodiversity and Ecosystem Services. IPBES Secretariat. [Google Scholar]
- Wickham, H. , Averick M., Bryan J., et al. 2019. “Welcome to the Tidyverse.”
- Xu, L. , Biswas A., Fang F., and Tambe M.. 2022. “Ranked Prioritization of Groups in Combinatorial Bandit Allocation. arXiv.” 10.48550/arXiv.2205.05659. [DOI]
- Zhang, A. , Wang X., Gu X., Xu X., Gao X., and Jiao L.. 2025. “Estimation of the Aboveground Carbon Stocks Based on Tree Species Identification in Saihanba Plantation Forest.” Ecological Indicators 173: 113370. 10.1016/j.ecolind.2025.113370. [DOI] [Google Scholar]
- Zizka, A. , Andermann T., and Silvestro D.. 2022. “Iucnn–Deep Learning Approaches to Approximate Species' Extinction Risk.” Diversity and Distributions 28, no. 2: 227–241. 10.1111/ddi.13450. [DOI] [Google Scholar]
- Zu Ermgassen, S. O. S. E. , and Löfqvist S.. 2024. “Financing Ecosystem Restoration.” Current Biology 34, no. 9: R412–R417. 10.1016/j.cub.2024.02.031. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Table S1: Biodiversity credit valuation estimated to reach the same net financial gain as in Policy 1 (focused on carbon sequestration) across all other policies (with increasing weight associated with extinction risks), computed based on Equation (17). The values are calculated across 100 simulations based on the reference region (on which models were trained).
Table S2: Biodiversity credit valuation estimated to reach the same net financial gain as in Policy 1 across all policies, computed based on Equation (17). The values are calculated across 100 simulations based on a set of random regions differing in size and location from the region used to train the policies (Figure S3).
Table S3: Summary of the outcomes of 100 simulations performed under different restoration models. Nonparametric models 1–3 (see Supplementary Methods and Results: Appendix S1) were based on simple metrics targeting carbon and costs (1); carbon, biodiversity, and costs (2); and biodiversity and costs (3). The other models (Policies 1–5) were based on CAPTAIN‐trained neural networks of 3 and 2 nodes in two hidden layers (small NN) and 16 and 8 nodes (larger NN). The effectively protected species quantify the number of species that started from a threatened category (CR, EN, and VU) at the beginning of the simulations but moved to a non‐threatened category (NT and LC) after restoration, while the species with improved status are species that moved to a lower extinction risk. All values are given as median among simulations and 2.5% and 97.5% quantiles rounded to integer.
Figure S1: Changes in spatial restoration priorities depending on the relative weights of carbon (highest in Policy 1) and biodiversity (highest in Policy 5). (a) Example of the restored areas in one of the 100 simulations performed on the reference region used to optimize the models (Figure S2) selected under five policies (Table 2). Note that the placement of the restored areas depends on the specific conditions of the simulated dataset. Restored areas will therefore vary among simulations, reflecting the randomly initialized distribution of species, disturbance, and costs. (b) Inferred restoration priorities obtained from simulations based on another, randomly selected region, showing similar discrepancies in spatial prioritization depending on the relative importance attributed to carbon versus biodiversity outcomes (Figure S3).
Figure S2: Map of the arbitrary area used for model training and subsequent simulations with examples of the empirical and simulated data used in our CAPTAIN runs. The physical properties of the area are based on the real topography and climate of the selected portion of Central and South America (a). The area was gridded into 128 × 128 cells and included empirical climate and elevation layers, while the “natural” state of biodiversity and carbon (b) and the disturbance patterns and costs (c) were simulated and changed randomly across simulations. Species richness and carbon maps are obtained after stacking 500 simulated species (see examples in d). Note that only the map in panel (a) is based on real data, while all other panels show simulated data. Similar simulations were repeated on 100 additional random maps of different size and location (Figure S3).
Figure S3: Maps showing the 100 randomly sampled rasters used as the basis for our simulations. The raster highlighted in red indicates the region used to train the models and to run the simulations based on a fixed geographic extent (see Figure S2 and Section 2).
Figure S4: Training rewards across epochs based on a small policy network (see Section 2).
Figure S5: Training rewards across epochs based on a larger policy network (see Section 2).
Figure S6: Summary statistics of key outcomes from 100 simulations run under five trained restoration models, Policies 1–5. The net carbon (a) represents the net gain in amount of carbon (here expressed in arbitrary units—see Section 2) in the system after 30 time‐steps during which 3000 cells were restored. As the weight assigned to the carbon reward was lowered from 0.5 to 0 in favor of the biodiversity reward, the net carbon gain gradually decreased, although all restoration policies resulted in a net gain. In contrast, the number of species with improved conservation status, that is, lower extinction risk, drastically increased as soon as a non‐zero weight was assigned to the biodiversity reward (Policies 2–5; b). The implementation cost was similar across all policies (c), reflecting the fact that all models were trained with the same weight on the budget reward. Plots (d–h) show the difference between the number of species in each extinction risk category, going from highest risk (CR) to lowest (LC), at the beginning of the simulation and the corresponding number at the end of the simulation. Thus, negative numbers, here observed in (d)–(g) mean that the number of species in the corresponding risk categories decreased. Positive numbers, here mostly observed in (h), mean that the number of species in these categories increased. These plots show that all policies that considered biodiversity rewards in the optimization (Policies 2–5) resulted in substantial reductions of threatened species compared to a policy targeting carbon only. This is also evident when lumping together all threatened species (i).
Figure S7: Summary of 100 simulations from random locations and variable resolution (Figure S3) run under five trained restoration models, Policies 1–5. The general trends in terms of net carbon gain (decreasing from Policy 1 to 5; a), number of species with improved conservation status (higher from Policies 2–5; b), and costs (roughly constant across all policies) are similar to those observed across the simulations based on a fixed geographic extent, upon which the models were optimized. However the patterns are less consistent when looking at changes in species number across risk category (from CR to LC; d–h) or moving from a threatened to a non‐threatened classification (i). These results suggest that the models (here trained on a single region) could be re‐optimized or fine‐tuned to best fit different regions.
Figure S8: Financial outcomes of restoration policies (net percentage loss or gain, that is, in Equation (15)) under different valuations of biodiversity credits, obtained from running the simulations across 100 random regions with different spatial location, extent and resolution. Negative values indicate a net economic loss where restoration costs exceed the cumulative value of carbon and biodiversity credits, while positive values indicate a net gain. The results are shown when only carbon credits () are included (i.e., ; a), which we used as a baseline of net 0 financial outcome to show how biodiversity credits could contribute value to the simulated restoration programs. The results in (b)–(d) show that when biodiversity credits are valued at , all policies become financially viable ().
Figure S9: Restoration priorities in the Atlantic Forest: results for the three policies show the frequency at which each grid cell was selected for restoration (green‐blue shades) across 30 replicates, where growth rates, species sensitivity to disturbance, and missing extinction risk assessments were randomly imputed (a–c). Sensitivity tests show the impact of simulated error in per‐species carbon (d–f) and in the spatial distribution of restoration costs (g–i).
Figure S10: Correlation coefficients for the seven climatic variables (Fick and Hijmans 2017) retained for species distribution modeling.
Figure S11: Violin plots and box plots of the five performance metrics of species distribution models across (a) all species and (b) retained species after applying exclusion criteria (sensitivity < 0.7, specificity < 0.7, True Skill Statistic < 0.5, and Area Under Curve < 0.8).
Figure S12: Approximate values of (a) restoration costs, (b) anthropogenic disturbance, (c) simulated total carbon in the absence of disturbance, and (d) simulated species richness based on stacking the species distribution models. The values are approximated as described in the Section 2.
Figure S13: Example of Perlin noise applied as random cost multiplier to evaluate the impact of error in the cost layer on the inferred restoration priorities. The multiplier is rescaled to alter the cost layer by ±25%.
Data Availability Statement
All the code and data used to reproduce the results shown in this study are available on Dryad: https://doi.org/10.5061/dryad.r7sqv9svg. The latest version of the CAPTAIN software is available here: https://www.captain‐project.net.

