SHAP-Driven Temporal Transformers for Transparent Crop Yield Prediction

Ensuring accurate and interpretable crop yield predictions is vital for addressing food security issues and guiding agricultural policies as climate change intensifies. We present Cultivating Clarity, an innovative framework that combines a Temporal Transformer with SHAP (SHapley Additive exPlanations) to deliver reliable and clear yield forecasts. By using Area embeddings to capture spatial patterns and analyzing five-year sequences of climate, pesticide, and temperature data, our model builds resilience with$\pm 20 \%$climate variations, inspired by IPCC reports. A Random Forest baseline leads with an$R^{2}$of 0.9746 without augmentation, while our Transformer achieves$\boldsymbol{R}^{2}$values of 0.9564 (without) and 0.9473 (with), excelling in tracking temporal trends and handling multimodal data. SHAP highlights global and local factors, like rainfall and temperature impacting yams in Colombia and maize in Armenia. An ablation study shows a minor accuracy dip with augmentation due to noise, striking a balance between toughness and precision.