Mapping agricultural dynamics in displacement-affected landscapes using satellite foundation model embeddings

Abstract

Sub-Saharan Africa already faces low human development, extreme climate risks, and environmental degradation–pressures compounded by large-scale population displacement resulting from armed conflict. Nowhere are these intersecting challenges more evident than in Uganda, which currently hosts the largest refugee population in Africa (nearly 2 million), straining land management, natural resources, and local food systems. As is common across displacement settings, spatially explicit baseline agricultural data in Uganda’s refugee-hosting regions remain scarce, limiting land-use planning, food security assessments, and humanitarian interventions. We present the first high-resolution assessment of cropland dynamics across refugee settlements in northern Uganda (2017–2024) using satellite foundation model embeddings and machine learning. We leveraged Google’s 10 m AlphaEarth Foundation embeddings with Random Forest classification to generate annual cropland maps, and applied sample-based accuracy assessment and area estimation to quantify uncertainty. Overall classification accuracy ranged from 0.75 to 0.87 (F1-scores: 0.65–0.84). Area-adjusted estimates revealed pronounced interannual variability: rapid expansion from 2017 to 2019, contraction in 2020 coinciding with COVID-19 disruptions and flooding events, a peak extent in 2022 (447,809 ±12,733 ha), and a marked decline by 2024 (223,159 ±9,744 ha) coinciding with humanitarian aid funding cuts and drought events. To quantify drivers of these dynamics, we fit a linear mixed-effects model relating cropland cover to distance from settlement boundaries, rainfall and temperature anomalies, and year. Cropland cover declined significantly with distance from settlement boundaries ( β=−3.08, p<0.001) and was positively associated with rainfall anomalies ( β=3.37, p<0.001), while temperature anomaly was not a significant predictor. These results provide quantitative evidence that settlement proximity and rainfall deficits are associated with observed cropland dynamics, though the observational design does not establish causality. By integrating foundation-model representations with rigorous uncertainty quantification and driver analysis, we provide settlement-scale evidence to inform land-use planning, food security assessments, and humanitarian interventions in fragile, displacement-affected landscapes.