A Hybrid Spatio-Temporal Transformer-LSTM Network for Explainable Crop Yield Prediction Using Multi-Source Remote Sensing in Arid African Regions

Abstract

The escalating food security crisis in Sub-Saharan Africa, driven by population growth and climate volatility, necessitates robust agricultural monitoring and prediction systems. However, existing approaches often rely on historical trends and struggle to capture the nonlinear effects of erratic rainfall and thermal shocks, particularly in arid regions. Furthermore, limited geographic representation and opaque black-box algorithms can constrain the use of predictive models for agricultural decision-making. To address these limitations, this study proposes the Hybrid Spatio-Temporal Transformer-LSTM Network (HST-TLN), which integrates multi-source environmental data, including Sentinel-2 multispectral imagery, ERA5-Land meteorological data, Sentinel-5P atmospheric nitrogen dioxide (NO₂), soil variables, and historical yield information. The architecture employs a dual-pathway mechanism in which Long Short-Term Memory (LSTM) units capture sequential temporal dependencies, while a Multi-Head Self-Attention Transformer identifies important temporal patterns associated with environmental stress. An Adaptive Cross-Attention Gating Mechanism is further used to integrate the two representations, while Shapley Additive Explanations (SHAP) provides interpretable feature contributions. Experimental results show that HST-TLN achieves an R² of 0.94 in normal years and 0.89 during drought years, outperforming the evaluated benchmark models, with the largest improvement observed under climatic stress. Cross-regional evaluation further demonstrates its predictive capability in geographically unseen regions. These findings indicate that combining temporal memory and contextual attention can improve crop yield prediction under environmental volatility and support data-driven agricultural planning in developing regions.