Abstract
Climate variability increasingly disrupts agricultural productivity, demanding systems that couple high-volume data with interpretable AI. We present AgriClimateAI, a big-data analytics framework that unifies multi-source inputs— satellite imagery, meteorological records, and crop yield statistics—to monitor climate impacts and support decision making. At its core is ClimaCropNet, a CNN–LSTM hybrid that learns spatial patterns from remote-sensing features and temporal dependencies from climate trajectories, followed by an adaptive fusion layer to model climate–crop interactions jointly. To ensure transparency, AgriClimateAI integrates SHAP and LIME for global and local explanation, revealing key drivers and validating alignment with agronomic knowledge. Evaluated across multiple agro-climatic zones, ClimaCropNet achieved an R² of 0.85 and RMSE of 0.43 t/ha for yield forecasting, and 88.6% accuracy for climaterisk classification, consistently outperforming baseline machine learning and single-stream deep models. Explainability analyses ranked rainfall and NDVI as the most influential predictors, with consistent seasonal saliency across regions. The framework’s cloud-scalable design supports near real-time ingestion, spatiotemporal analytics, and deployment over diverse cropping systems and climates. By delivering accurate forecasts with auditable rationale, AgriClimateAI enables climate-smart advisories, adaptive input planning, and policy dashboards for resilient agriculture. Overall, ClimaCropNet advances interpretable spatiotemporal learning for integrated yield prediction and risk assessment, while AgriClimateAI operationalizes these capabilities into an end-to-end, transferable system for data-driven agricultural resilience.
Keywords: Climate-Smart Agriculture, Crop Yield Prediction, Explainable AI, Remote Sensing, Spatiotemporal Deep Learning.