Deeply divergent human exposure to food crises across socioeconomic pathways

Abstract

Hundreds of millions of people worldwide suffer from acute food insecurity. This emergency calls for timely preventive and response actions, which depend on our ability to anticipate food crises. Since food security results from complex combinations of societal, economic, political, and environmental factors, most existing models require extensive data and are based on many assumptions, limiting their realism and applicability across time and space. Here, we show that accurate models predicting the yearly onset of food crises can be generated using only temperature and precipitation data. When combined with demographic and poverty projections under different socioeconomic pathways, these models allow us to explore future global scenarios. Conflict and inequality pathways (SSP3 and SSP4) could expose more than 1.1 billion people—mostly in Africa and Asia—to at least one severe food crisis by century’s end. Of these, more than 600 millions would be under five years old at first exposure, and more than 230 millions would face a crisis in their first year of life. In contrast, a shift toward environmental and social sustainability (SSP1) could more than halve current yearly exposure and reduce worst-case cumulative exposure by 69%. These findings highlight that today’s policy decisions may lead to radically different food security futures.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-29068-1.

Introduction

Climate change affects societies and economies in multiple ways¹, playing a fundamental role in the functioning and stability of food systems^2–6. Some of the direct connections between climate and food security are evident. For instance, droughts can affect food availability and prices through detrimental impacts on agricultural production^7–11. More elusive are the indirect connections where climate modulates the effects of other drivers of food security, for example by promoting conflict^12,13, political instability¹⁴, and migration¹⁵. Extreme weather events and adverse conditions can increase human mortality^15–18, disrupt transportation¹⁹, and fuel the spread of disease and epidemics^20,21. These processes can influence local and global food availability through multiple, convoluted pathways of causal links and feedback mechanisms^2,4,6,22. While reconstructing all of these pathways with a theory-driven approach would be challenging, a data-driven approach might be able to exploit the information they underlie to devise effective predictive models^5,23,24. Here we test this hypothesis by exploring whether machine-learning models based on climate variables alone can predict the onset of acute food insecurity crises.

We obtained food security data from 2010 to 2022 (with three assessments per year) from the Famine Early Warning System Network (FEWS NET)²⁵. Specifically, we used the data referring to the current food-security assessments provided by FEWS NET. These represent evidence-based evaluations of the prevailing situation, in contrast to the FEWS NET projections, which are forward-looking forecasts combining observed data with modelling approaches and expert analysis²⁶. These data come in the form of five alert levels (from 1 “minimal”, to 5 “famine”) for countries in Central America and the Caribbean, Central Asia, East Africa, Southern Africa and West Africa, which we projected on a regular grid at a resolution of 0.5 × 0.5 latitude/longitude degrees, as in²⁷. We then identified the onset of a food crisis (hereafter, simply a “food crisis”) in a grid cell, when the conditions changed from level 1-2 to a level ≥3 and remained like that for at least two consecutive assessments, as in²⁸. This criterion captures specific but fundamental facets of food security, and points to emergency situations that call for immediate response²⁸. The total number of new food crises recorded per year showed a substantial increment (almost doubling) in the period 2011-2020 (Fig. 1a), with the total number of people exposed to food insecurity levels ≥3 almost tripling (from a little more than 50 to almost 150M) (Fig. 1b).

Fig. 1.

The number of people exposed to food crises almost tripled within the past decade. The left panel (a) shows the total number of people that faced the onset of a new food crisis per year in the areas monitored by FEWS NET (https://fews.net/²⁵), which is what our model aims to predict. There, the figures for a given year do not include the people facing a continued crisis from the previous year(s). Conversely, the right panel (b) shows the number of people that faced at least one period of food alert level ≥3 in a given year, regardless of the food security state of previous assessment periods.

We combined this information with meteorological data^29,30 to devise a machine-learning (random forest³¹) model capable of predicting whether a locality would experience at least one food crisis within a target year. As predictors, we used monthly average temperature and precipitation for the 24 months preceding the target year, along with derivative variables derived from these series (see Methods). This choice reflects the assumption that meteorological patterns influence multiple direct and indirect drivers of food security, and that the combined temperature–precipitation time series may act as characteristic “fingerprints” of food-insecurity triggers.

The model showed a high overall accuracy, with average over- and underestimation rates of 0.19±0.05 and 0.13±0.06 respectively, and average True Skills Statistics of 0.68±0.06 SD (assessed through spatio-temporal cross validation, n=100; see Methods for details). The total number of predicted crises per year and per country showed a high correlation with those observed in the cross validation (R²=0.65; see Supplementary Fig. S1). A test performed on the Global Report on Food Crises database³² (which includes a list of major food crises and associated drivers recorded at country level during the period 2017-2024), showed that false predictions occurred at a higher frequency when the main putative cause of crisis was conflict (61.0% and 59.3% of all false positives and negatives respectively) compared to settings in which crises were attributed to weather extremes (21.2% and 23.9%) or economic shocks (17.7% and 16.8%). Error rates did not show any strong trend across years (Supplementary Fig. S2).

Overall, these results demonstrate that temperature and precipitation alone provide an effective indicator for the onset of yearly food crises, confirming the lead role played by climate change in food security. Precipitation variables had a stronger effect on the model’s predictions compared to temperature, with a precipitation deficit (especially in winter—December to February) emerging, not surprisingly, as a major determinant of (modelled) food crises (see Supplementary Figs. S3-S6).

We then applied the model to future climate projections (from the 6th phase of the Coupled Model Intercomparison Project, CMIP6³³) at a spatial resolution of 1 × 1 degree latitude/longitude (i.e. the highest resolution available for multi-model ensemble projections), and limiting the projections within the latitudinal range 40S-40N (approximately the latitudinal extent currently covered by localities monitored by FEWS NET). We explored four different economic pathways depicting alternative futures where the world either: shifts towards a more sustainable path within environmental boundaries, posing low challenges to mitigation and to adaptation (SSP1, “Sustainability - Taking the green road”; associated with a radiative forcing of 2.6 W/m²); does not shift substantially from historical social, economic and technological patterns, with moderate mitigation and adaptation challenges (SSP2, “Middle of the road”, associated with a radiative forcing of 4.5 W/m²); moves towards international fragmentation and regional rivalry, resulting in high mitigation and adaptation challenges (SSP3, “ Regional rivalry - A rocky road”, associated with a radiative forcing of 7 W/m²); transitions towards deep inequality between wealthy, technologically advanced and well educated societies internationally connected; and fragmented, poorly educated, lower income societies, with low challenges to mitigation, but high challenges to adaptation (SSP4, “Inequality - A road divided”, associated with a radiative forcing of 6 W/m²)^34,35.

To obtain cautious estimates of future risk, we used a conservative probability threshold for random forest classification, aimed at keeping overestimation error rates below 0.05 (see Methods). Even with this adjustment, the model predicts a consistent and steady increase in the occurrence of climatic conditions associated with food crises in the coming decades under different socioeconomic pathways³⁵ (Fig. 2A; Supplementary Table S1). However, this does not necessarily translate into an actual risk, as crisis-triggering conditions could materialise in spatio-temporal settings that are either underpopulated, or are made food-secure by the local socioeconomic system. We therefore applied an additional conservative criterion based on the documented relationship between income and different aspects of food security^48,49. Specifically, we combined the model predictions with demographic³⁶ and economic projections³⁷ (see also Supplementary Fig. S5), and we assumed as food secure all localities with a per-capita gross domestic product above the 95^th quantile of the values where a crisis was recorded in the FEWS NET dataset (i.e. 9,333 US dollars in 2005). The model predicts that the future number of people exposed to food crisis risk will vary wildly depending on the socioeconomic pathway (Fig. 2b).

Fig. 2.

Contrasting effects of different socioeconomic pathways on the future of global food security. (a) Probability of food crisis onset predicted by the model alone (i.e. without considering local population and GDP), averaged across all localities, quantifying the climate component of food crisis risk. (b) Number of people predicted to be exposed to a food crisis per year, computed as the sum of people in all localities where the model predicted the occurrence of a food crisis, without considering local GDP. (c) Number of people predicted to be exposed a food crisis per year, computed as the sum of people in all localities where the model predicted the occurrence of a food crisis and where the projected GDP per capita was lower than the GDP per capita threshold (9,333 US dollars in 2005). In panel a, shaded areas indicate standard error of the mean probability, while in panels b and c they indicate 95% confidence intervals of predictions. To ease visualization, a 5-y moving average was applied to all curves.

Pathways moving towards sustainability substantially reduce the projected number of exposed people per year by mid- and end of the century. Specifically, SSP1-2.6 leads to a relative 75% (log % difference) decrease in the average yearly human exposure in the last decade of the century compared to 2005-2015 (from 89.2±9.5M to 42.3M±13.8, mean±SD, see Supplementary Table S1). Conversely, future conflicts and inequality increase the projected number of exposed people to 280.4±42.3M (under SSP3-7.0) and 229.1±65.8M (under SSP4-6.0). Compared to the most optimistic scenario (SSP1-2.6), the conflict and inequality scenarios might lead, respectively, to 6.6× and 5.4× more endangered people in 2090-2099. Even transitioning from sustainability (SSP1) to mid-road (SSP2) scenarios dramatically changes the projections, more than doubling the numbers of people exposed (2.1×).

Different assumptions regarding the effects of GDP, as well as different choices regarding the probability threshold for random forest predictions could lead to different numbers depicting more or less pessimistic projections. However, the substantially higher estimates of SSP3 and SSP4 with respect to SSP1 and 2 persists across a wide range of assumptions (Supplementary Fig. S7). The differences between SSP4 and SSP1 exceed several billions of exposed people under the most pessimistic assumptions (Supplementary Fig. S8a,c), and the projections of human exposure in SSP4 are on average 246% greater than those in SSP1 across all parametrizations (Supplementary Fig. S8b,d).

We combined model predictions with country-level demographic projections (consistent with the different socioeconomic pathways) to compute the cumulative number of people predicted to experience at least one famine crisis during the century, accounting for future births and deaths (Fig 3; Supplementary Table S2). Under the inequality pathway (SSP4), that number exceeds 1.16 billion. As observed for the annual risk, the differences between the expected levels of exposure under different scenarios are huge. Potentially, SSP1 could spare 784M people from risk by 2099 (Fig 3a, dashed line). The majority of exposed people will be children under 5 years of age at the time of the first crisis experienced (631M and 620M under SSP3 and SSP4 respectively, Fig. 3b), and hundreds of millions of newborns will face a crisis within their first year (266M under SSP3, and 231M under SSP4 Fig. 3c).

Fig. 3.

Cumulative exposure to food crisis risk. (a) Cumulative number of people of all ages (in billions) that will experience at least one food crisis by the end of the century under different shared socioeconomic pathways. (b) Cumulative number of children younger than 5 (in billions) that will experience at least one food crisis by the end of the century. (c) Cumulative number of newborns (in billions) that will experience at least one food crisis within their first year by the end of the century. The coloured continuous lines represent sums of exposed people across all localities per year. The black dashed lines represent the difference between the inequality socioeconomic pathway (SSP4-6.0) and the sustainability one (SSP1-2.6). In all panels, shaded areas indicate 95% confidence intervals of predictions.

The risk will be disproportionate across continents (Figs. 4,5; Supplementary Table S3). Even though the model identified potential risk in areas currently not monitored by FEWS NET¹¹ (Fig. 4), it predicted most of the future crises in regions coinciding with or close to monitored areas (Figs. 4-5). Consistent with the expectation that climate change will exacerbate food insecurity in areas which are already insecure⁴, most future crises will be in Africa or Asia. Although crises will be much more widespread in Africa in terms of area (Fig. 4), projections in terms of people exposure (Fig. 5) are made comparable between the two continents by the higher population density in Asia. However, although under SSP3 and SSP4 the model projects a steady, almost linear and parallel increase in population exposure for both continents, the trajectories show a substantial divergence under more optimistic scenarios. Especially under SSP1, we predict a rapid decline in yearly exposure after mid-century for Africa, while the situation in Asia is expected to remain close to the initial state (with similar but less pronounced patterns for SSP2).

Fig. 4.

Mapping the future risk of food crisis under different shared socioeconomic pathways. Maps show the number of people exposed per year (in millions, with logarithmic colour scale), averaged over the two periods 2017-2050 (left panels) and 2051-2099 (right panels). The area shown by the maps (40S-40N, 25W-95E) covers 98.5% of total exposed population (across all scenarios and years).

Fig. 5.

Contrasting effects of different socioeconomic pathways on the future of global food security across continents. Plots report the number of people predicted to be exposed to food crises per year, computed as the sum of people in all localities where the model predicted the occurrence of a food crisis and where the projected GDP per capita was lower than the GDP per capita threshold (9333 US dollars referred to 2005). Shaded areas indicate 95% confidence intervals of predictions. To ease visualization, a 5-y moving average was applied to all curves.

Conclusions

The model’s ability to identify food crises with high accuracy at a medium-grained time and spatial resolution further supports the importance of climate in global patterns of food security^2,4,6,22. Although the climate’s role develops through a combination of direct and indirect, convoluted causal pathways, our findings demonstrate that lack of knowledge about the intermediate nodes that link climate to food security should not prevent us from attempting to explain observed patterns or to make future projections based on available information. This does not imply that our approach should replace more sophisticated models that account for a broader spectrum of social, economic and politic factors, including the current, near-term predictions provided by FEWS NET. By contrast, our model could serve as a complement to them. Its main value and novelty lie, in fact, in its ability to project future food security risk based on readily available data. This makes it a promising tool to increase preparedness by mapping the risk of food crises over a substantially longer time horizon than permitted by other available models (such as, for instance,²⁸). Even more, this makes the approach particularly suitable to perform scenario analyses, and examine the potential consequences of climate mitigation policies, as we did here. Model projections could also be used as a frame of reference to pinpoint and investigate real-world cases where factors other than climate had a larger (or smaller) than expected effect on food security. For instance, focusing on mismatches between predicted and observed crises might help better understand and quantify the role of socio-political insecurity on food crises, as supported by our finding that the majority of the model’s false predictions are associated with events of conflict.

Our projections align with other model-based predictions on other dimensions of food security. Specifically, they are consistent with other studies that identify massive differences between different SSPs^38,39, and that project numbers of people at risk by the end of the century within the same order of magnitude (for instance, varying among scenarios from a few million to 600 M in³⁸).

At the continent level, the differences between the scenarios are not just in terms of magnitude of human exposure, but also in terms of quality of the predicted trajectories. Under SSP3 and SSP4, Africa and Asia show consistent, rapidly growing patterns of exposure. By contrast, under SSP2 and (even more) SSP1, the trajectories will diverge substantially, with the exposure in Asia remaining stable, while that in Africa showing a marked decline. These results highlight that even sustainable socioeconomic futures might not translate into significant improvements in Asia’s food security (at least for the dimensions covered by our metric).

One important caveat of our results is that our computation of exposure based on local population and GDP underlies the simplified assumption of uniform vulnerability across social classes. However, considering global patterns in the distribution of wealth and inequality (especially in low- and middle-income countries)^40–42, the assumption should have a negligible impact on the magnitude of our estimates. Actually, considering typical distributions of household income and wealth⁴³, it is reasonable to assume that our simplification is conservative, as it could lead to overestimating the safety of a large portion of populations in areas where GDP pro-capita is close to the selected threshold. Increasing that threshold could lead to predictions of substantially greater numbers of people at risk (Supplementary Figs. S7, S8).

The model depends on the accuracy of FEWS NET in recording and assessing food crises. There, biases might emerge from potential issues in data quality and availability on famine events, mostly due to political factors resulting, for instance, in reluctance to provide accurate records on mortality⁴⁴. In general, however, such biases are more likely to result in under-reporting (rather than over-reporting) crises, which might partly explain the overestimation errors of the model⁴⁵.

Another caveat to our projections is that they do not take into account both potential future improvements in the local resilience and response ability of local stakeholders and administrations, as well as the potential beneficial effects of humanitarian actions put in place by international and intergovernmental organisations, which could mitigate future emergencies^46–48. Similarly, they do not consider potential future transformations in food systems driven by technological progress, even though this is, to at least to some extent, an important part of the narrative in the SSP scenarios³⁴. Still, the consistency in the extensive gaps between the predicted outcomes of alternative socioeconomic pathways on global food security under a large combination of assumptions and model parameters (see Figs. 2-5 and Supplementary Figs. S7, S8) questions the feasibility of putting in place effective responses to future scenarios that deviate even slightly from sustainability. In addition, we should consider that our projections refer to one specific dimension of food security. The onset of food crises is particularly important for preparedness, as it helps pinpoint emergencies that call for immediate intervention. Still, by measuring only a single facet of food security we could neglect a large share of people at risk, including, for instance, people exposed to chronic or continuous crises (Fig. 1).

Our results confirm that global food security is now at a crucial crossroads, and that following one or another pathway will lead to completely different futures, affecting the lives of hundreds of millions of people³⁸. Although the current societal and economic momentum make it difficult to steer away from inequality and towards sustainability⁴⁹, our findings further highlight that failing to do so will not only have devastating impacts on the natural world, but might also threaten billions of human lives^50–54.

Methods

Identifying food crisis

We obtained food security data from the Famine Early Warning System Network (FEWS NET, https://fews.net/)²⁵. FEWS NET provides standardized food insecurity severity data for multiple countries in Central America and the Caribbean, Central Asia, East Africa, Southern Africa, and West Africa from 2010 to present days, with three assessments per year²⁵. As its core product, FEWS NET provides food-security forecasts aimed at improving preparedness by anticipating potential critical situations. However, alongside with forecasts, it also provides data-driven assessments of the “current” food-insecurity at a given moment. These can be used a posteriori to evaluate the accuracy of future projections⁴⁵ a or to train models²⁸. FEWS NET quantifies food insecurity severity using a discrete scale spanning five distinct phases (compatible with the “Integrated Food Security Phase Classification”, IPC), from “generally food secure” (level 1) to “famine” (level 5). Following the procedure in²⁷, we interpolated all FEWS NET data on “current” food insecurity on a regular grid at a resolution of 0.5 × 0.5 latitude/longitude degrees. We then identified the onset of a food crisis (hereafter “food crisis”) in a grid cell at a given assessment time t whenever the food security phase in that cell changed from level 1 or 2 (minimal or stressed) at time t-1 to ≥3 (crisis, emergency or famine) at time t, and remained like that at least in the next period, t+1 (as in²⁸). We then aggregated data at yearly intervals, considering a year of crisis any year where at least one crisis happened. We limited the dataset to only grid cells in areas monitored by FEWS NET, so to avoid inflating false absences by including areas not monitored. In total, we obtained a dataset of 76608 observations across 6125 localities (grid cells) over 13 years and including 7178 food crises.

Modelling food crises

We obtained meteorological data from publicly available datasets. Specifically we obtained average monthly temperature from NOAA (https://psl.noaa.gov/data/gridded/)²⁹ and precipitation from CHIRPS (https://www.chc.ucsb.edu/data/chirps)⁵⁵, that we matched with the food crisis data at a spatial resolution of 0.5 x 0.5 degrees. Using these data, we associated to each food crisis value (0 or 1) in a given location at a given year 48 values indicating the monthly temperature and precipitation in the 2 years preceding the target one. We also added 10 composite yearly variables obtained from monthly values, namely: cumulative yearly temperature/precipitation (i.e. sum of monthly values); standard deviation of monthly temperature/precipitation; maximum and minimum temperature/precipitation; temperature/precipitation range (difference between maximum and minimum monthly values).

We then used the dataset to train a random forest classifier³¹ aiming at predicting the occurrence of food crisis events in a given locality based on these independent variables. To evaluate model accuracy while accounting for potential spatio-temporal autocorrelation in the data, we performed a two-fold cross validation^56,57 where we generated multiple spatially and temporally independent pairs of training and testing sets. To this end, we first explored how the dependency between crisis observations changed with their pairwise geographical distance using join counts statistics⁵⁸.

For this, we generated 100 random subsamples of unique localities from the dataset, by drawing recursively random localities one at a time and including them only if they were at least a given threshold distance from all previously selected sites. Threshold distances were varied from 50 to 150 km in 10 km increments, and from 200 to 500 km in 50 km increments. For each subsample, the test measured the deviation (Z score) between the observed and expected frequency of presence–presence joins among neighboring localities. The resulting Z-scores declined rapidly with distance and became statistically negligible (Z < 1.96, p>0.05) beyond ~120 km, which we therefore considered the minimum separation distance required to ensure approximate spatial independence in food-crisis risk between neighboring localities (see Supplementary Figure S9).

Based on this criterion, we generated 100 pairs of spatially independent training and testing sets. For each pair, we first sampled at random 1% of all the localities in the complete dataset, and allocated all the records in the dataset falling in those localities to the testing set (discarding testing sets having less than 50 presences). We then allocated to the training set all the records from the remaining localities at at least 120 km from the closest point in the testing set.

Note that this procedure not only implies full spatial segregation between the training and the testing set, but also temporal separation. The spatial separation is imposed by the distance buffer between any point in the testing set and the points in the training set; while temporal segregation is ensured by the fact that all observations from the same location (across all years) belong to the same split, and therefore training and validation are never performed on the same site at a different time. That is, the procedure cross-validation design ensures that no point from the training set has future (or past) occurrences in the testing set.

We used the training sets to calibrate 100 random forest models⁵⁹, and we assessed their performance by comparing their predictions with the observed values in the testing sets, quantifying accuracy in terms of false positive and negative error rates, and true skills statistics ⁶⁰. For each model we explored a continuous range of threshold values for class attribution probability, varying it in between 0.001 and 0.9 (with increments of 0.001)^61–63. We then retained the threshold leading to the higher TSS score for the accuracy assessment. For each model, we also recorded the smallest probability threshold keeping false positive error rate below 0.05, and we used the average of those thresholds to derive (conservative) future projections.

We generated the random forest models using the R package ranger⁶⁴ . Each model included 1000 trees, and we set the number of random variables to be tried at each split to 7 (that is, to the square root of the number of independent variables, rounded down, which is the standard choice in random forest modelling; note that we tested alternative parametrizations with no appreciable improvement in model accuracy). To cope with the high class-imbalance of the dataset, where 90.6% of records are identified as non-crises (0s) and only 9.4% are identified as crises (1s), in addition to adjusting the classification threshold⁶¹, we used a weighted design in the random forest implementation⁶⁵ where the contribution of observations to the model was proportional to class prevalence. For that, we set the “case.weights” variable in ranger’s⁶⁴ random forest function to 0.094 for non-crisis records, and to 0.906 for crisis records. We finally trained a model on the complete dataset, with the same set-up used in the cross-validation. We used that complete model for all the analyses and projections. We quantified variable importance in terms of mean impurity decrease.

To better understand the model’s mechanism, we explored the marginal effects of individual variables on the predicted probability of crisis⁶⁶. For this, we first derived a reference probability by fitting the model on the average values of all variables in the FEWS NET dataset. Then we explored how that probability changed keeping all other variables fixed to their mean value, while varying the variable of interest over a broad range of values, and specifically: 0-400mm for precipitation; 260-320K for temperature; minimum-maximum observed value from the training dataset for the composite variables. We divided each range into 1000 equal steps. We quantified marginal effect as the log percentage difference between the reference probability and the probabilities over the range of variation of the variable of interest. The range of marginal effect values for a given variable provided additional information on variable importance⁶⁶.

Sensitivity analysis

Using the same procedure used for the cross-validation of the main model, we also assessed the performance of models based on temperature and precipitation data spanning either 12 or 36 months before the target crisis. Both models showed a performance comparable to the 24-month model (see Supplementary Table S4). Furthermore, as an additional test to ensure robustness of our model against spatial-autocorrelation, we performed an alternative spatial cross validation (for models using a time window of 24 months for the definition of the independent variables) by training the random forest classifier on a random sample including 99% of FEWS NET administrative areas (https://fews.net/data/geographic-boundaries), and testing it on the remaining areas. The cross validation yielded an average accuracy comparable to that obtained using the 120km buffer (see Supplementary Table S4).

We also performed an additional test to explore the potential effects of temporal auto-correlation on the model’s predictive ability, by constraining the training set to records in the years 2010-2019, and then testing the model on records from 2021-2022 (and not falling within 80km of distance from any point in the training set). There, the cross-validation accuracy resulted still high, despite the exercise reduced the potential training data by 22% and, more importantly, lowered the number of recorded crisis to be possibly used in training by 24% (Supplementary Table S4).

Evaluating model’s accuracy for different drivers of food crises

Data on the main driver of food crisis from 2017 to 2024 are available from the Food Security Information Network (FSIN)’s Global Report on Food Crises database (GRFCd available at https://www.fsinplatform.org/report/global-report-food-crises-2024)⁶⁷. The database includes a list of major food crisis recorded at country level in a given year, and the associated main putative driver (weather extreme, economic shock, conflict). We included this information in the cross validation procedure to evaluate the model’s predictive ability under different food-insecurity scenarios. Specifically, for each one of the three main drivers, we quantified under and over estimation errors using all occurrences from the gridded dataset falling in a given country and year for which a major crisis attributed to the target driver was included in GRFCd. We then quantified the percentage of over and under estimation errors falling in settings (locality × year) associated to a specific driver. Note that the exercise was applicable not only to false negatives, but also to false positive, due to the subnational resolution of FEWS NET spatial data. Hence, a false positive would consist of a cell in the 0.5 x 0.5 degree grid where the model predicts a crisis not reported in FEWS NET data, but falling within a country/year where GRFCd reported a crisis and the corresponding driver.

Projecting future food crises

We obtained future monthly and precipitation multi-model ensemble projections (from 2020 to 2099) for four future socio economic pathways (sustainability, SSP1-2.6; middle of the road, SSP2-4.5; conflict, SSP3-7.0; inequality, SSP4-6.0) at the global scale and at a spatial resolution of 1×1 latitude/longitude degrees (the best resolution available for multi-model ensembles at monthly intervals) from⁶⁸. The ensembles are based on, respectively, 26, 27, 22, and 5 models for SSP1-2.6, SSP2-4,5, SSP3-7.0, SSP4-6.0 (see⁶⁸ for details). To ensure consistency between the (real-world) climate data used to train the model, and the future (modelled) projections, we performed a delta-change bias correction⁷⁵. For each SSP scenario, we computed absolute temperature monthly anomalies in respect to the average values in the historical CMIP6 data for the period 1981-2010, and we then summed those values to the corresponding observed averages (from NOAA temperature) for the same reference period⁶⁹. To correct future precipitation data, we used percentage change, computing their relative anomalies in respect to the reference values from the historical scenario, and then multiplying the anomalies by the average of observed precipitation in the reference period (from CHIRPS data).

We fed these data to our model to obtain yearly, global scale projections of potential food crises from 2022 to 2099. We limited the projections within the latitudinal range 40S-40N. Such projections identify locations (1×1 degree grid cells) where the meteorological conditions of the two years preceding the target one are favourable to the onset of a food crisis. As a conservative choice for the probability threshold for classification, we took the average of the maximum thresholds recorded in each one of the 100 models in the cross-validation procedure keeping the overestimation error rates below 0.05. We also computed 95% exact binomial confidence intervals for the random forest predictions (according to Pearson-Klopper method⁷⁰) based on the binary classifications yielded by each one of the 1000 individual trees of the model. For that, we use the R package binom⁷¹.

The risk predicted by the model does not necessarily translate into an actual risk in terms of human exposure to food insecurity, as the socioeconomic context might mitigate the climate-driven hazard. For this, we obtained gridded future GDP projections (at 5 year intervals, from 2020 to 2100) consistent with the socio economic pathways from³⁷. To take into account inflation (and country-level inflation differences) we transformed all values (for both the historical data and the future data) into their corresponding (country-specific) 2005 US dollar⁷² value, using the R package priceR⁷³. We then limited predicted future food crises to locations with projected GDP pro capita below the 95^th percentile of the corresponding value recorded at the time and location of a crisis reported in the FEWS NET dataset (9,333 US dollars referred to 2005).

An obvious alternative approach to setting a GDP-based threshold for the prediction of crises would be that of including GDP into the model as an additional independent variable. However, such an approach would be suboptimal due to the fact that the distribution of GDP in the areas covered by the model is far from being representative of global distributions, and characterized by a relatively small variability compared to the full range of values observed outside FEWS NET zones, which would substantially compromise the model’s extrapolation ability (see supplementary Fig. S10).

We obtained gridded future projections of population density (again, consistent with the selected socio economic pathways) from³⁶. The original dataset has a resolution of 30 arc seconds, which we rescaled using bilinear interpolation to match our 1×1 degree grid. For current/past population density (2000-2023), we obtained data from LandScan global dataset (https://landscan.ornl.gov/)^74–96. Again, we interpolated the data (which were at an original resolution of 30 arc-seconds) on our 1×1 degree grid. We then used these data to quantify food security risk in terms of total number of exposed individuals per grid cell, summarizing the data at both global and continent level. As for the GDP data, ideally one could have included population density as an additional independent variable in the model. However, that would have been problematic, being a potential source of circularity in the quantification of population exposure, with the projected risk being partially driven by local population density, and then further combined with the same variable to obtain the final value of human exposure.

To explore how and to what extent the projected human exposure to crises is affected by more or less conservative choices in the random forest classification probability and GDP thresholds, we performed extensive sensitivity analyses. In particular, we generated three large sets of alternative future projections for each SSP scenario. In the first set, we kept fixed the GDP per-capita threshold (9,333 US dollars referred to 2005), while we varied the random forest probability threshold in (0.100, 0.105, …, 0.500). In the second set, we kept fixed the probability threshold (to the same value used for the future projections, see Methods for details), varying the GDP per-capita threshold in (100, 200, …, 20000). In the third set, we varied both the random forest probability and the GDP thresholds, in the same intervals as above, computing human exposure at 2099.

Quantifying the cumulative number of exposed people

We obtained demographic projections consistent with the selected socio economic pathways from⁹⁷. Specifically, we used the “low” and “medium” fertility scenarios for SSP1-2.6 and SSP2-4.5, and the “high” fertility scenario for SSP3-7.0 and SSP4-6.0⁹⁸. These data are at country level and at a yearly time resolution. They provide age class population distribution using 5 year classes from 0 to 100+ years. We converted those to 1 year classes by generating a smooth piecewise-quadratic probability density function based on the binned age classes, using the splinebins function of the binsmooth package⁹⁹. Then, we projected the country level values to the 1×1 degree grid cells, by multiplying the values of the probability density function over (0, 1, …, 100) for each country/year per the total population in the grid cells corresponding to that country/year. This resulted in the estimated number of individuals for each yearly age class from 0 to 100 in each grid cell. We then quantified the cumulative number of individuals exposed to at least one crisis across the period 2022-2099 by summing up, for each year of crisis occurring at time t_i in a given location, the number of individuals belonging to age classes <t_i-t_i-1. To quantify the number of children less than 5 who experienced at least one food crisis across the period 2022-2099, we summed up, for each year of crisis occurring at time t_i in a given location, the number of individuals belonging to age classes <5 and <t_i-t_i-1. Finally, we quantified the number of newborns exposed to a crisis within their first year by summing up the number of individuals of age 0 in each cell/year hit by a crisis.

Author contributions

G.S. designed the research, performed the analyses and wrote the paper.

Data availability

Code and data permitting full replication of the analyses are available with no restrictions from the Zenodo repository https://zenodo.org/records/17177371.

Declarations

Competing interests

The authors declare no competing interests.