The Importance of Landscape Composition for Pest Control and Crop Yield: A Global Quantitative Synthesis

ABSTRACT

Land‐use change has altered the composition of our landscapes to favour agriculture, negatively affecting biodiversity and ecosystem services. However, the links between landscape composition, pest control and yield remain unclear. Using a global structural equation model of 116 studies from 28 countries, we tested three hypotheses: the ‘natural enemy hypothesis’, that natural areas increase natural enemies and suppress pests; the ‘resource concentration hypothesis’, that more agriculture increases pests; and the ‘agronomic quality hypothesis’, that agriculture‐dominated landscapes occur in high‐yielding areas. Results show landscape composition affects yield directly and indirectly through crop, herbivore and natural enemy traits. At larger scales (~1250 m), natural habitats increase yield, but at smaller scales (~250 m), yields decline with more surrounding natural habitat, likely due to lower agronomic quality or edge effects. Our findings suggest that we can harness the positive effects of natural areas at large scales while mitigating drawbacks at smaller scales to promote sustainable production.

In a global analysis of 116 studies, we show that landscape composition affects crop yield both directly and indirectly via traits of crops, herbivores and natural enemies. Our results suggest that we can harness the positive effects of large‐scale natural areas while potential yield trade‐offs at smaller scales should be carefully managed.

1. Introduction

Land‐use change is one of the main drivers of global biodiversity loss (Newbold et al. 2015; Sala 2000), with cascading effects on ecosystem services that underpin agricultural productivity and farmers' livelihoods (Bianchi et al. 2006; Ricketts et al. 2008). These changes impede sustainable intensification, the challenge of increasing food production while minimizing environmental impacts (Pretty and Bharucha 2014; Rockström et al. 2009). One of the key strategies for achieving this balance is integrating landscape ecology principles into agricultural systems; specifically, maintaining or restoring landscape heterogeneity as a means to enhance biodiversity and ecosystem services (Kremen and Merenlender 2018; Tscharntke et al. 2012). However, economic pressures often necessitate that farmers prioritize short‐term productivity over long‐term sustainability, risking the degradation of ecosystem services essential for sustained agricultural yields into the future (Giller et al. 2009; Pittelkow et al. 2015; Seppelt et al. 2020).

Landscape simplification, defined as the conversion of complex landscapes rich in natural habitats to vast, homogenous agricultural fields, has been shown to reduce ecosystem services (Bianchi et al. 2006; Chaplin‐Kramer et al. 2011; Kennedy et al. 2013; Rusch et al. 2016; Veres et al. 2013), increase reliance on external inputs such as pesticides and fertilizers (Huseth et al. 2015; Meehan et al. 2011; Paredes et al. 2021), and, paradoxically, may even decrease yields (Dainese et al. 2019). To address these challenges, several policies worldwide promote practices to enhance landscape heterogeneity, aiming to boost biodiversity while simultaneously supporting ecosystem services like crop yield. Policies promote landscape heterogeneity as a tool for sustainable intensification (Tscharntke et al. 2021), with the aim of boosting biodiversity and ecosystem services, ultimately resulting in higher crop yields. Nonetheless, previous global data syntheses examining the relationship between landscape composition and crop production have reached differing conclusions. Karp et al. (2018) reported no consistent effects of landscape composition on pest control or yields, while Dainese et al. (2019) identified a potential yield decline in agriculture‐dominated landscapes. These seemingly inconsistent results could be attributed to the fact that Karp et al. (2018) analyzed the direct effects of landscape composition on natural enemies, herbivores and yield but did not explicitly consider the indirect pathways through which natural enemies and herbivores influence agricultural productivity. In contrast, Dainese et al. (2019) employed a structural equation modeling (SEM) approach to disentangle both direct and indirect effects, demonstrating that biodiversity loss in agriculture‐dominated landscapes can reduce crop yields via declines in pollinators and natural enemies. However, neither study explicitly accounted for the indirect effect of natural enemies and herbivores in shaping the relationships between landscape composition and yield, leaving a critical gap in our understanding of the direct and indirect mechanisms by which landscape composition impacts agricultural sustainability. Here, we extend previous analyses by employing a structural equation modeling approach that incorporates herbivore dynamics, predator–prey interactions and crop traits, allowing us to disentangle multiple ecological pathways influencing agricultural productivity.

To date, two hypotheses have been proposed to explain how landscape composition affects yield, mediated by herbivores. The natural enemy hypothesis (Figure 1A), initially proposed at local (i.e., field or farm) scales (Root 1973), is now commonly used at the landscape scale (O'Rourke and Petersen 2017). It proposes that landscapes dominated by (semi‐) natural habitat maintain high natural enemy abundance and diversity, which in turn provide effective biological control (Bianchi et al. 2006; Chaplin‐Kramer et al. 2011, 2013; Dainese et al. 2019; Werling et al. 2011). This increased biological control has been hypothesized to lead to reduced herbivore abundance, reduced pest pressure (O'Rourke and Petersen 2017; Tscharntke et al. 2016) and, subsequently, increased yield. The resource concentration hypothesis (Figure 1B), also initially proposed at local scales (Root 1973), posits that fields dominated by a single crop type offer more resources to specialised herbivore pests, leading to higher plant damage and reduced yield. This hypothesis has been extended to the landscape scale, where studies suggest that simple landscapes dominated by monocultures may provide pests with concentrated resources that allow populations to rapidly build, disperse and inflict significant crop damage (Bianchi et al. 2006; O'Rourke and Petersen 2017; Thies et al. 2003). These effects are expected to be especially pronounced for specialist herbivores, while generalist herbivores that feed on a variety of agricultural and natural plants (Santoiemma et al. 2018) would benefit more from increased landscape heterogeneity.

FIGURE 1.

Three (non‐mutually exclusive) hypotheses (A–C) for how landscape complexity may influence yield directly through potential changes in agronomic conditions or indirectly through changes in natural enemy and/or herbivore communities. (A) Natural enemy hypothesis, (B) Resource concentration hypothesis and (C) Agronomic quality hypothesis. In addition, traits from herbivores (D), natural enemies (E) and cropping systems (F) might alter the direction or strength of these effects. Black arrows denote positive effects while red arrows denote negative effects. Grey dashed arrows show other potential links not tested in the presented hypothesis or scenario. The thickness of the arrow denotes the strength of the hypothesized relationship.

Along with these well‐recognised hypotheses that involve indirect effects of herbivores on yield, landscape composition in itself could be an indicator of agronomic quality. Agricultural land is preferentially established in areas of higher agronomic quality (Bürgi and Turner 2002; Neyret et al. 2023; Serneels and Lambin 2001; Verburg et al. 2004), such as those with higher soil fertility, favourable climate and easier water access. As a result, areas with ideal biophysical characteristics for farming should be dominated by agricultural land uses, resulting in a positive association between yield and agriculturally dominated landscapes. For this reason, we here propose the agronomic quality hypothesis (Figure 1C), which suggests that crop yields (i.e., production per area) will peak in simplified landscapes, dominated by agriculture. The degree to which each hypothesis is reflected in real‐world landscapes may depend on the functional traits of both the crops and their interactors, which determine species' responses to land use change and their contributions to ecosystem services (Duflot et al. 2014; Gámez‐Virués et al. 2015; Martin et al. 2019; Tamburini et al. 2020). Moreover, the functional trait composition of a community can affect the outcome of plant‐insect interactions, including crop production. For example, specialist but not generalist herbivores are known to respond to resource concentration at both local (Root 1973) and landscape (O'Rourke and Petersen 2017) scales. In addition, specialist herbivores can overcome plant defences more effectively than generalists and therefore inflict more damage on well‐defended plants (Bidart‐Bouzat and Kliebenstein 2008; Mithen et al. 1995). Crop traits may also play an important role in the outcome of plant‐herbivore interactions. Crops relying on outcrossing and pollinators to reach optimal yields may be disproportionately affected by land use changes, given the negative effect that conversion of land to agriculture has on pollinators (Dainese et al. 2019; Kennedy et al. 2013). Natural enemy traits also influence pest control efficacy. Although parasitoids, with their more specialised host selection, are often thought to provide more effective biological control (Van Veen et al. 2008), generalist predators may exert stronger top‐down control (Jiang and Morin 2005; Karp et al. 2013; Symondson et al. 2002) due to their broad prey range (Karp et al. 2013) and resilience in human‐dominated ecosystems such as agriculturally dominated landscapes (Corbett et al. 2024; Tscharntke, Bommarco, et al. 2008).

Here, we conduct a quantitative synthesis of 116 studies examining how landscape composition affects natural enemies, pests and crop yields. Specifically, we explored the direct and indirect effects of landscape complexity on yield based on previously hypothesized mechanisms (Figure 1) resulting in the following predictions: (1) Landscapes dominated by (semi‐) natural habitats improve yields through an increased natural enemy richness and abundance that leads to a reduction in pests (Figure 1A). (2) Specialist herbivores increase in landscapes dominated by agriculture through a concentration of resources, leading to more plant damage and a reduction in yield (Figure 1B). (3) Specialist herbivores should, at the same time, have a stronger negative effect on crop yield than generalist herbivores (Figure 1C). (4) Landscapes dominated by agriculture should have higher yields due to their inherently high agronomic quality (Figure 1D). (5) Predators should contribute more to biological control than parasitoids, as they are better adapted to human‐dominated landscapes (Figure 1E). (6) Pollinator‐dependent crops should show higher yield with increasing adjacent (semi‐) natural habitats, compared to pollinator‐independent crops (Figure 1F). To test these predictions, we use structural equation models (SEMs), which are designed to estimate direct and indirect effects among multiple variables. This approach allows us to disentangle the direct effect of landscape composition on yield from the indirect effects mediated through natural enemies and/or herbivores, while considering traits of crops (pollinator dependence), herbivores (specialisation) and natural enemies (parasitoid vs. predator).

2. Materials and Methods

2.1. Data Source

The data in this paper were collected, analysed and published in collaboration with Karp et al. (2018). The resulting global database includes published and unpublished studies gathered from 85 data contributors who provided data on natural enemy and pest abundances, crop damage and/or pest predation/parasitism rates in crop fields across at least six sites along a landscape gradient (mean number of sites per study = 62.3, range = 6–2108). Each data contributor filled out a standardised data entry form that was sent to them. For our analysis, we used 116 studies (i.e., a subset of the original dataset published in (Karp et al. 2018)). Specifically for this study, we excluded all records that (1) focused on non‐arthropod taxa (i.e., birds or bats), (2) lacked insect abundance or biological control data, (3) reported ‘yield’ in non‐crop systems (e.g., grasslands) and (4) did not specify the crop sampled. In some instances, we reclassified abundance data from the original dataset as biological control information; for example, when parasitism rates were reported as parasitoid abundance. Abundance was defined as the number of individuals per plant, trap or square meter (e.g., aphids per plant or spiders per pitfall trap).

To determine the importance of pest specialisation, we quantified total pest abundance as well as the abundance of specialist and generalist pests, based on (Tamburini et al. 2020) and additional knowledge obtained from published literature, extension sites and internet searches for taxa not included in their study (see Supporting Information S1 for full classification). We classified herbivores as ‘specialist’ when they feed on plants from a single family, whereas ‘generalist’ included herbivores that feed on plants belonging to more than one family (following (Martin et al. 2019)). Species with unknown host preferences were classified as ‘unknown’. Species classified only by genus, were denoted ‘specialists’ if the genus is entirely composed of specialists, ‘generalist’ if entirely composed of generalist species and ‘unknown’ if composed of a mix of species. Natural enemy data were categorised into total natural enemy abundance, and then divided into parasitoids and predators.

Biological control included any measure of the impact of a natural enemy on a pest, excluding abundance data. This included exclusion experiments, sentinel prey items removed by predators and parasitism rates. Yield data covered both marketable and total yield (see Supporting Information S2 for details). We manually evaluated each study to ensure proper links between enemy abundance, biological control, pest abundance and yield. For example, studies reporting both parasitism and predation rates were split into separate observations (study subsets with different study ID) in order to correctly link predator abundance to predation and parasitoid abundance to parasitism. This evaluation increased our sample size from 116 studies reported in Karp et al. 2018 to 127 studies.

2.2. Data Standardisation

Given the variability in data collection methods, and species with different baseline abundances, (e.g., number of caterpillars per plant ranged in the tens, while aphids ranged in the thousands), we standardised our data as in Karp et al. 2018, per sampling method, year and study. All response variables were standardised to a common unit of collection effort; specifically, if sampling occurred over a varying number of days or number of traps, all data from a given study were adjusted by the number of census types (i.e., traps or transects) and by duration (e.g., total collected pests per trap per day). Subsequently, the data were normalised to a zero mean and unit variance within each sampling method, year and study. This ensured comparable responses across sampling methods, years and studies.

2.3. Landscape Composition

In Karp et al. (2018) we quantified landscape composition within 2 km of each site and in the same timeframe (year(s)) the study was performed. The land cover types included: (1) natural and semi‐natural areas (forest, non‐crop tree plantations, grassland and scrubland) hereafter referred to as ‘natural’, and (2) cropping areas (annual cropland, and perennial cropland) hereafter referred to as ‘agricultural’. Landscape composition was quantified using the same hierarchical approach as in Karp et al. 2018. In short, the proportion of land cover types were obtained from high‐resolution land‐use maps provided by the authors, regional maps available online (e.g., the National Land Cover Database, NLCD) or, when necessary, a global land cover product (https://www.un‐spider.org/links‐and‐resources/data‐sources/land‐cover‐map‐globeland‐30‐ngcc). The composition of each cover type around each site was quantified within concentric rings of 100 m, with closer rings weighted more heavily using a Gaussian decay function (decay rates: 250, 750 and 1250; see Karp et al. 2018). To ensure sufficient variation along the landscape composition gradient and adequate characterisation of the surrounding land cover, studies were excluded if < 75% of the surrounding landscape was quantified or if both agricultural and natural land cover showed < 10% variation across sites.

2.4. Hypothesis Testing

To evaluate how landscape complexity could directly or indirectly affect crop yield as proposed in our hypotheses (Figure 1), we adopted a two‐step information‐theoretic approach (Jørgensen 2004). First, we selected the variables that best explained our hypothesized relationships (Supporting Information S3), and second, we constructed piecewise structural equation models (SEM) (Shipley 2009) using those selected variables. Year nested within a study was included as a random effect in all models.

2.5. Factor Selection

Before constructing SEM, we used mixed effect models (nlme v 3.1‐168, Pinheiro et al. 2025) to select the most predictive variables. Because authors measured yield in two different ways, either as marketable or total yield, we first, tested whether the type of yield measure affected the yield value. Finding no effect, we excluded yield type from the SEM analysis. Second, we built individual additive mixed‐effect models for all response variables including all landscape metrics (Supporting Information S3). We used the dredge function from the MuMIn package in R (version 1.48.11, Barton 2025) to rank models by the lowest corrected Akaike Information Criterion (AICc) value using maximum likelihood (Jørgensen 2004), retaining just landscape variables from the top‐ranked models (AIC < 2). Variables were transformed as needed to meet assumptions of normality and heteroskedasticity (Karp et al. 2018) (Supporting Information S3). During this initial screening, only the proportion of natural habitat at 250 and 1250 m were significant landscape predictors of any response variables and therefore retained for subsequent models. However, to test the resource concentration hypothesis (i.e., how the proportion of a focal crop might affect specialist herbivores), we included the proportion of agriculture at 250 and 1250 m in all analyses. All other variables were kept for further analysis in the SEM.

2.6. Confirmatory Path Analysis: Direct and Indirect Effects on Yield

We constructed our initial structural equation model using the sem.fit function from the piecewiseSEM package (v 2.3.1, Lefcheck 2016) in R, combining individual mixed‐effects models into a piecewise SEM framework. Model structure was informed by previously identified key variables (Supporting Information S3) and designed to test the first three hypotheses outlined in Figure 1. The model included yield as the primary response variable, with pest abundance, natural enemy abundance and the proportion of natural and agricultural habitat at two spatial scales (250 and 1250 m) as key predictors. The path from agricultural habitat proportion to yield was included to test the agronomic quality hypothesis (Figure 1C). To test the resource concentration hypothesis (Figure 1B) we modelled total pest abundance as a function of landscape composition. Pest abundance, in turn, was used as a predictor of yield. To evaluate the natural enemy hypothesis (Figure 1A), we modelled pest abundance as a function of total natural enemy abundance and biological control, while modelling these natural enemy metrics themselves as a function of landscape composition. To account for non‐independence across time, we included the year nested within the study as a random effect in all models. To test the conditional independence claims implied by the path models, we used Shipley's d‐separation test. Each claim assumes that a pair of variables should be statistically independent given the other variables in the model. Each claim is tested with linear mixed models, and the resulting p‐values are combined into a Fisher's chi‐squared distributed C‐statistic (Shipley 2009). A non‐significant C (p > 0.1) indicates adequate model fit (Shipley 2013). To improve model fit, we modified our initial model using a backward and forward stepwise process first based on the significance of Fisher's C‐statistics and then on Akaike's information criterion (AIC). We started simplification with backward simplification of non‐significant exogenous variables (landscape metrics), followed by non‐significant endogenous variables. Missing paths proposed by the d‐separation test, which are paths required to satisfy the conditional independence assumption implied by the data, were then added to the model, making sure that the Fisher's test of fit had a p > 0.1 (Shipley 2013). When individual paths were still not significant, they were individually removed, and the model with the lowest AIC was maintained, which sometimes included models with non‐significant paths. Model assumptions were evaluated by identifying outliers, testing for multicollinearity using variance inflation factors (VIFs, from the ‘car’ package version 3.1‐3 in R, Fox and Weisberg 2019), and assessing spatial autocorrelation using Moran's I (‘spdep’ package version 1.4‐1, Bivand and Wong 2018) (see Supporting Information S6 for results). No evidence of problematic multicollinearity was found (all VIFs < 3.6). However, significant spatial autocorrelation was detected in a subset of models (Supporting Information S6). Although we attempted to incorporate spatial error structures within the piecewise SEM framework, these adjustments did not account for the spatial autocorrelation and the original models were retained. Outlier diagnostics revealed that most outliers were part of the expected data distribution; however, in the pollinator‐dependent model, a clear outlier group was identified and removed prior to reanalyzing the dataset. Path coefficients were calculated using the piecewiseSEM package (v 2.3.1, Lefcheck 2016). All analyses were conducted using R version 4.4.2 (R Core Team 2024).

2.7. Confirmatory Path Analysis: Functional Traits

In order to test each of the hypotheses related to organism traits (as described in Figure 1), we created four additional models in which (model 1) we replaced total enemy abundance with both the abundance of predators and parasitoids as separate variables, (model 2) we replaced total herbivore abundance with both the abundance of specialist and generalist herbivores as separate variables, and (models 3 and 4) we divided all crops into pollinator‐dependent versus independent crops. To perform the analyses on pollinator dependence, we divided our dataset into two, given that the herbivore and natural enemy communities were unique to each crop species, and we did not want to include the herbivore and natural enemy communities of the pollinator‐dependent crops in the analysis of the pollinator‐independent crops. Therefore, one dataset included all wind‐ or self‐pollinated crops (n = 76 studies), while the other dataset included all crops where there is evidence that pollinators can contribute to crop production (n = 47 studies) (For crop classification, see Supporting Information S5). For each dataset, we repeated the same analytical methods as the one described above using total pest and total enemy abundance. Although we tried to test all the functional traits simultaneously in one model, the model did not converge, due to insufficient replicates relative to the number of variables, hindering more detailed analyses.

3. Results

3.1. Database Characteristics

We obtained a total of 116 datasets (referred to as studies) from all five continents and 28 countries, with the majority originating from the United States (27 studies), followed by France (12 studies) and Germany (11 studies). The most frequently studied crop was wheat ( Triticum aestivum L.), investigated in 34 studies, followed by Brassica oleracea L. (including cabbage, broccoli, kale, etc.; 14 studies), and coffee ( Coffea arabica L.; 12 studies). Across all studies, the percentage of natural area at a larger scale (with a 1250 m decay rate) ranged from 4% to 72%, with an average within‐study range of 45%. For additional details on the database characteristics, see Supporting Information S5. The proportion of agriculture and natural habitat in the landscape were negatively related to each other at all scales (r < −0.61, p < 0.001; Supporting Information S6).

3.2. Overall Analysis: Direct and Indirect Effects on Yield (Figure 2A)

FIGURE 2.

(A) Structural equation model (SEM) showing direct and indirect effects of landscape complexity (proportion of natural habitat at 250 and 1250 m decay) on crop yields mediated through changes in the abundance of natural enemies, biological control and pest abundance. (B) SEM incorporating the natural enemy guilds by substituting the total abundance of natural enemies with separate abundance values for predators and parasitoids. (C) SEM incorporating the pest host specialisation by substituting the total pest abundance with separate abundance values for specialist and generalist pests. Black arrows represent significant positive effects, while red arrows represent significant negative effects. The light grey arrows represent the correlated error between landscape metrics. Numbers in the arrows are z‐standardised path correlation coefficients, which are interpreted as the expected change in standard deviation unit in the dependent variable for a one standard deviation increase in the predictor variable. The accompanying asterisk denotes their p‐value significance level (***p < 0.001, **p < 0.01, *p < 0.05, ˙ p < 0.10). AIC, chi‐squared and Fisher's C (test of fit) values with their respective p values and degrees of freedom are presented adjacent to each analysis. For more details about model structure and coefficients, see Supporting Information S6.

In all tested models, there was no support for including the relationship (direct or indirect) between the proportion of agriculture in the landscape and natural enemies, pests or yield and is therefore not depicted in any of the results (Figure 2). In contrast, our results clearly showed that the proportion of natural area in the landscape had both direct and indirect negative effects on pests and yield (Figure 2A). An increase in the proportion of natural habitat at a smaller scale (250 m decay) had a direct negative effect on yield. Indirectly, natural areas at a larger scale (1250 m decay) affected pests, mediated by a positive effect on natural enemies. Specifically, a higher abundance of natural enemies was correlated with higher biological control, and higher biological control was correlated with lower pest abundance. Interestingly, pest abundance was not associated with yield.

3.3. Natural Enemy Traits (Figure 2B)

When we isolated the specific effects of individual enemy guilds, we found that the proportion of natural habitat at a smaller scale but not at a larger scale had an indirect effect on pests, mediated by parasitoids (Figure 2B). Specifically, parasitoids displayed a positive correlation with the proportion of natural habitat at 250 m, and parasitoids were negatively associated with pest abundance. Parasitoids were not directly linked to biological control (i.e., parasitism rates). In contrast, predators were associated with higher biological control (i.e., predation rates), which was associated with lower pest abundance. Despite this indirect negative effect on pests, predator abundance also exhibited a positive correlation with pest abundance. As in the overall analysis, an increase in natural habitats at 250 m was negatively correlated with yield.

3.4. Pest Traits (Figure 2C)

We found that specialist pests, but not generalist pests, mediate an indirect effect between landscape composition and crop yield (Figure 2C). A higher proportion of natural areas at a larger scale was negatively associated with specialist pests, which further had a negative correlation with crop yield. On the other hand, generalist pests responded indirectly to changes in the proportion of natural habitat at a larger scale. Specifically, an increase in natural areas at larger scales was associated with an increase in the abundance of natural enemies, which were positively correlated to biological control and, in turn, a reduction in the number of generalist pests.

3.5. Crop Traits (Figure 3A,B)

FIGURE 3.

Structural equation model (SEM) showing direct and indirect effects of landscape complexity (proportion of natural and agricultural habitat at 250 and 1250 m decay) on crop yields mediated through changes in the abundance of natural enemies, biological control and the abundance of pests for pollinator‐dependent (A) and pollinator‐independent (B) cropping systems. Black arrows represent significant positive effects, while red arrows represent significant negative effects. Dashed arrows and coefficients represent marginally significant effects (0.05 > p > 0.1). Light grey arrows represent the correlated error between landscape metrics. Numbers in the arrows are z‐standardised path correlation coefficients, which are interpreted as the expected change in standard deviation unit in the dependent variable for a one standard deviation increase in the predictor variable. The accompanying asterisk denotes their p‐value significance level (***p < 0.001, **p < 0.01, *p < 0.05, ˙ p < 0.10). AIC, chi‐square and Fisher's C (test of fit) values with their respective p values and degrees of freedom are presented adjacent to each analysis. For more details about model structure and coefficients, see Supporting Information S6.

The ecological interactions between landscape composition, predators, pests and biological control represented in the path analyses were different for pollinator‐dependent and pollinator‐independent crops (Figure 3). For pollinator‐dependent crops, there was a positive effect of the proportion of natural areas at a 1250 m decay scale on yield (Figure 3A). This positive effect was not mediated by natural enemies or pests, suggesting a potential role of pollinators. In contrast, the proportion of natural habitat at a 1250 m decay scale did not significantly affect pollinator‐independent crops directly or indirectly (Figure 3B).

Pollinator‐dependent crops (Figure 3A), but not pollinator‐independent crops, displayed a negative effect of natural habitat at a 250 m decay scale on yield, consistent with the overall analysis. The effects of natural habitat on enemies and pests mirrored the overall model for pollinator‐independent crops, where natural habitat at the largest scale was linked to increases in enemies, subsequently leading to enhanced biological control (marginally) and reduced pest abundance (Figure 3B). For pollinator‐dependent crops, there was a positive association between natural enemies and pests, which may reflect a density‐dependent response of natural enemies to pest abundance. Interestingly, we also observed a positive effect of natural areas and agricultural areas at a larger scale on pest abundance (Figure 3A). Given that the subset of data for pollinator‐dependent crops lacked sufficient information on biological control data, we could not determine its role in this model.

4. Discussion

This study demonstrates the direct and indirect impact of landscape composition on crop yield. By employing structural equation models in conjunction with species traits, we have refined our understanding of these effects, particularly in relation to herbivorous insects and their natural enemies. This approach has enabled us to uncover nuanced interactions that eluded previous analytical methods (Karp et al. 2018). Our findings offer partial support for the natural enemy and agronomic quality hypotheses, and indirectly also for the resource concentration hypothesis. Notably, the magnitude and direction of these effects frequently hinged on factors such as the spatial scale of natural habitats and the specific traits of the organisms and crops involved.

4.1. The Natural Enemy Hypothesis

The natural enemy hypothesis was partially supported by our data. An increase in the proportion of natural habitat at larger scales (1250 m decay) was associated with higher natural enemy abundance, which increased biological control and reduced pests. However, the overall path analysis did not indicate a direct negative effect of pests on yield, supporting other studies that have suggested natural enemy effects on pest abundance do not always translate into increased yields (Letourneau et al. 2011; Poveda et al. 2008). This may occur when pest populations remain below economic damage thresholds (Losey and Vaughan 2006), or due to compensatory plant responses that mitigate damage (Poveda et al. 2018; Strauss et al. 1999). We acknowledge that several important ecological and agronomic mechanisms, such as intraguild predation (Rosenheim and Harmon 2006), crop variety selection (Evenson and Gollin 2003), fertilisation regimes (Ludwig et al. 2011), as well as other on‐farm management practices, could influence both pest dynamics and yield outcomes. While our study focused on landscape‐level patterns, future work that integrates field‐level management data and biotic interactions could help disentangle the multiple, potentially confounding drivers of pest–yield relationships.

Interestingly, these indirect links between the proportion of natural habitat at larger scales and pests, mediated by natural enemies and biological control, held for generalist pests and for pollinator‐independent crops, but not for specialist pests and pollinator‐dependent crops. This demonstrates the importance of pest and crop traits in influencing these interactions, as we discuss below.

Although overall we found that natural areas at a larger scale increased natural enemy abundance, the pattern was less clear when we separated natural enemies into parasitoids and predators. First, the effect of natural habitats was now significant at a smaller scale and not at a larger scale. Second, only parasitoids showed a positive relationship with the proportion of natural habitats, whereas predators were not affected by natural habitats at all. This scale‐dependent shift may reflect differences in dispersal capacity and resource dependence between enemy groups. Parasitoids are often more specialized and have limited dispersal ability (e.g., < 1 km), making them more sensitive to the immediate landscape context and the availability of floral and host resources in nearby natural habitats (Kruess and Tscharntke 2000; Thies and Tscharntke 1999). In contrast, generalist predators such as spiders, carabid beetles and ants tend to be more mobile, less dependent on specific habitat features and better able to persist in simplified landscapes or recolonize after disturbance, making them more resilient to changes in agricultural systems (Chaplin‐Kramer et al. 2011; Lindell et al. 2004; Tscharntke, Sekercioglu, et al. 2008). Moreover, differences in foraging behavior and trophic interactions could also explain the distinct scale responses. Parasitoids often track herbivore hosts closely and require specific microhabitats, while predators may exploit a broader prey base and operate across larger spatial and temporal scales (Bianchi et al. 2006). Interestingly, when we analyzed the data to investigate the effect mediated by parasitoids vs. predators on pest control, we see that both play a role in suppressing pests. Parasitoid abundance had a direct negative effect on pest abundance, while the effect of predators was mediated by measures of biological control. Interestingly, the direct link between predator and pest abundance was positive, probably due to density‐dependent effects, where increases in pests lead to increases in predators (Donaldson et al. 2007). This finding may also reflect intraguild predation or interference among predator species, which can weaken top‐down control despite high predator densities (Rosenheim et al. 1995; Finke and Denno 2004).

4.2. The Resource Concentration Hypothesis: Specialist Versus Generalist Herbivores

Our data partially supported predictions from the resource concentration hypothesis. The resource concentration hypothesis, initially proposed by Root 1973, was designed to explain the behaviour of specialist pests at local scales. According to this hypothesis, specialist pests that rely on particular plant species, or a limited group of plant species, should respond positively when those plants are more abundant (more concentrated). Interestingly, we find evidence for this hypothesis for pollinator‐dependent plants and specialist pests, but not in any of our other analyses. The fact that natural land and agricultural land directly affected pests in two cases may have to do with our inability to directly quantify the host plants of highest importance to specialist pests. Indeed, natural lands and agricultural lands can be highly variable and, even for specialists, certain agricultural or natural habitats can harbour more resources than others (Paredes et al. 2021; Priyadarshana et al. 2024; Rosenheim et al. 2022). Variations in management, such as insecticide use and crop rotation can also play important roles in the accessibility of resources. We suggest that future studies should assess the abundance of a certain plant family in the landscape (regardless of whether it is an agricultural or natural habitat) to properly assess the resource concentration hypothesis at a landscape scale.

Interestingly, our analysis revealed the importance of pest specialisation for yield. We found that specialist pests drove negative effects on yield, while generalist pests did not. There is evidence showing that specialist insect herbivores are capable of finding, eating and causing more damage to plants than generalist herbivore species (Mithen et al. 1995; Bernays and Chapman 2007; Bidart‐Bouzat and Kliebenstein 2008), making it clear why insect specialisation may play such an important role for crop yield.

4.3. The Agronomic Quality Hypothesis

We hypothesized that lands better suited for crop production and that, consequently have been intensified are more frequently used for agriculture worldwide. Our findings support this, showing that natural habitats at a small scale have a negative impact on yield in our overall model and for pollinator‐dependent crops. This pattern supports our hypothesis that agricultural fields near natural areas are often situated in areas of lower agronomic quality. However, we cannot exclude other alternative mechanisms. For example, areas with high natural habitat cover at small scales are also more likely to be located at field margins, also known as headlands. Recent evidence suggests that yields can be reduced by 14%–16% in these marginal areas in comparison to field centers due to the greater impacts from wildlife damage and soil compaction from machinery (Sunoj et al. 2021). Similarly, the ‘edge of field’ syndrome, which includes drought stress, competition and uneven fertilisation, may also explain the observed negative effects (Raatz et al. 2019; Robinson et al. 2022). The negative effect of natural areas at small scales on pollinator‐dependent, but not pollinator‐independent, crops may be explained by two non‐exclusive mechanisms. First, fruits and vegetables, which rely more on pollinators, are more likely to be planted on suboptimal land, that is of lower agronomic quality than grains (Bock et al. 2018). Second, a higher investment in intensive management practices (i.e., incorporating transgenic traits like Bt toxins and herbicide resistance) aimed mostly at grains (Soto‐Gómez and Pérez‐Rodríguez 2022), may mitigate the negative edge effects in pollinator‐independent plants.

4.4. Landscape Effects on Pollinator‐Dependent and Pollinator‐Independent Crops

We found different patterns between flowering crops that require insect pollination versus crops that do not require insect pollinators. In line with our predictions, we found a positive effect of natural areas at a 1250 m scale on pollinator‐dependent crops but not on pollinator‐independent crops. Given the ample evidence that landscapes with more natural areas have a higher pollinator richness, evenness and abundance (Dainese et al. 2019; Martin et al. 2019), we were not surprised to find that landscapes with more natural areas at a larger scale were associated with higher yields. Surprisingly, it was only in pollinator‐independent crop systems that we found support for the natural enemy hypothesis, with natural areas at the large scale increasing natural enemies and biological control and subsequently reducing pests. As with the overall analysis discussed above, there was still no evidence for an effect of reduced pest pressure on yield. In pollinator‐dependent cropping systems, we saw an increase in pest abundance in response to both natural and agricultural areas at larger spatial scales. This pattern aligns well with previous research on how landscape composition influences pest populations. There is substantial evidence that pests can benefit from both natural (Bianchi et al. 2006; Tscharntke et al. 2007) and agricultural (Meehan et al. 2011; Veres et al. 2013) habitats at the landscape scale, which often leads to ambiguous or inconsistent patterns in reviews and large scale analyses (Karp et al. 2018). The fact that we were able to disentangle these effects in our model and demonstrate that pests respond positively to both habitat types underscores the generality and pervasiveness of these patterns. As shown by Karp et al. (2018), such dual influences complicate the interpretation of landscape effects, particularly when habitat types are aggregated or not analysed independently. Future work should aim to disentangle which additional traits, beyond pest specialisation, may explain this pattern and to categorise the landscape more carefully to better predict when and why certain pests benefit from specific landscape types. This would enhance the development of management practices that take into account landscape complexity.

Overall, our findings provide three key insights. First, the complex and sometimes contradictory outcomes observed in landscape studies concerning natural enemies, pests and yield (Karp et al. 2018) can be partially attributed to the traits of the interacting organisms and the indirect effects mediated by these organisms. Traits such as pest diet breadth and natural enemy foraging strategy strongly modulate landscape effects, underscoring the value of trait‐based approaches in ecological intensification. Second, our results emphasize the need for management strategies that explicitly consider scale‐dependent processes. Large‐scale natural areas were consistently associated with positive effects on crucial ecosystem services, namely biological control, pest suppression and, very likely, pollination. However, at the small scale, natural areas appeared to be negatively correlated with yield, potentially due to lower agronomic quality at field edges. Although these findings may seem contradictory, they are encouraging in the sense that we continue to find a positive influence of large‐scale natural areas on ecosystem services, which, in some cases, can translate into positive yields. On the other hand, the negative effects observed at the small scale could be effectively managed using existing agronomic tools targeted to the edge of the field. While our analysis did not quantify the specific yield losses caused by these edge effects, addressing management practices such as fertilization, irrigation and plant density at the field edge may be sufficient to alleviate these costs. Third, our results emphasize that sustainable management must be tailored to specific cropping systems by integrating landscape context, organismal traits and local field practices. For instance, systems dominated by specialist pests or reliant on parasitoids may benefit from floral‐rich habitats at smaller scales, whereas broader‐scale conservation efforts could enhance generalist predator activity.

We are aware of two important caveats to our study. First, we recognize that we cannot fully rule out spatial dependencies in some of our models. Temporal and spatial autocorrelation are inherent challenges in ecological studies conducted across diverse landscapes and time periods. In our analysis, we detected significant spatial structure in a subset of models. While we attempted to account for this by including spatial error structures where possible, the piecewise SEM framework does not currently support fully flexible spatial modeling. Although spatial autocorrelation can inflate Type I errors and bias effect estimates, the consistency of observed patterns, such as the negative effects of natural areas at a small scale on yield, the positive effects of natural areas at the large scale on natural enemies or the negative effect of biological control on pests, suggests that our main conclusions are robust. Nevertheless, future studies would benefit from using spatially explicit modeling frameworks or hierarchical meta‐analytic approaches that more comprehensively address spatial and temporal structure.

Second, we recognize that this study does not offer system‐specific prescriptions, but provides a framework to guide management and future modeling efforts. Importantly, although our analysis identified associations between small‐scale natural habitats and reduced yield, it did not establish the direct causes of these reductions. Future research should prioritize uncovering the mechanisms driving these yield losses to better inform intervention strategies. Nonetheless, if we want to establish and preserve natural habitats within agricultural landscapes, a strategy that promotes biodiversity conservation and benefits yield, it becomes imperative to meticulously consider and address the prospect of yield losses when formulating subsidy programs. Designing subsidies or incentive programs that acknowledge both the benefits and drawbacks of landscape complexity will be key to promoting ecologically sound and economically viable agroecosystems.

In conclusion, our study highlights the importance of considering organismal traits and indirect interactions in landscape studies, as well as the need for scale‐dependent management approaches. By integrating these insights into agricultural practices, we can harness the positive effects of large‐scale natural areas while effectively managing the potential drawbacks at the small scale, ultimately promoting more sustainable and productive agroecosystems.

Author Contributions

Katja Poveda and Heather Grab conceived the ideas and designed the methodology. All authors collected the data. Heather Grab led the data analysis with help from Katja Poveda, Tim Luttermoser, Ricardo Perez‐Alvarez and Mary Centrella. Katja Poveda led the writing of the manuscript. All authors contributed critically and substantially to revisions and gave final approval for publication.

Supporting information

Data S1: Insect herbivore species (pests) included in this analysis and that were assigned a feeding guild based on Tamburini et al. (2020) and literature and internet searches. We considered an herbivore to be a ‘specialist’ when they feed on host plants belonging to just one plant family and a ‘generalist’ when they fed on plants belonging to more than one family. For some species we could not find any information on host plant preference and those were classified as ‘unknown’.

Data S2: Definition of response variables and the number of studies that reported these variables.

Data S3: Factor selection before the Structural Equation Model.

Data S4: List of crops used in this study, classified based on their pollinator dependence for crop production. This classification was mostly based on the data published by Klein et al. (2006), where crops were classified as (1) not increasing production in the presence of pollinators, (2) increasing production in the presence of pollinators, (3) increasing seed production in the presence of pollinators but not crop production and (4) pollinators being important for breeding purposes but not for productivity (Klein et al. 2006). Here, we determined crops to be pollinator‐dependent only if the presence of pollinators increased crop production, counting increases in seed production or importance for breeding (for plants with non‐reproductive yields, such as broccoli, cauliflower and cabbage) as non‐dependent on pollination. The list includes the reference from which we gathered the information on pollinator dependence (if found).

Data S5: Details on collected studies and dataset characteristics.

Data S6: Detailed interpretation of the structural equation model results.

Poveda, K. , Karp D. S., Chaplin‐Kramer R., et al. 2025. “The Importance of Landscape Composition for Pest Control and Crop Yield: A Global Quantitative Synthesis.” Ecology Letters 28, no. 11: e70250. 10.1111/ele.70250.

Data Availability Statement

All data and code used to analyse the data have been deposited in Dryad at: https://doi.org/10.5061/dryad.9ghx3ffwg.