Explainable Multimodal Artificial Intelligence Framework for Climate Resilient Crop Yield Prediction and Sustainable Farm Decision Sup-port

Abstract

Forecasting rice yield at the district level is central to agricultural planning and contingency management in India, yet the task is complicated by spatial heterogeneity, weather volatility, and strong temporal persistence in production. This study builds a provenance-audited, leakage-free predictive framework that pairs an XGBoost branch handling static and lagged agronomic predictors with a five-year Long Short-Term Memory (LSTM) network capturing sequential weather dynamics. Using a filtered multi-decadal panel of 300 Indian districts spanning 1981–2017, models were trained on data through 2012, tuned on 2013–2014, and independently evaluated on 703 chronologically held-out records from 2015–2017. A validation-selected late-fusion scheme (0.80 XGBoost, 0.20 LSTM) reached an RMSE of 477.29 kg/ha and an R2 of 0.743, meaningfully ahead of standalone XGBoost (485.87 kg/ha) and standalone LSTM (512.26 kg/ha). Paired statistical testing, however, found no significant edge over strong tabular baselines—Ridge Regression (477.44 kg/ha) and Random Forest (477.47 kg/ha). TreeSHAP attributions showed that predictions are driven mainly by historical yield persistence, secular year trends, and rainfall rather than intricate temporal patterns. Stratified diagnostics further revealed sharp performance loss under rainfall deficit (RMSE 547.71 kg/ha) and compound heat-dryness stress (560.93 kg/ha), together with pronounced disparities across macro-regions. Taken together, the results suggest that although hybrid architectures offer useful diagnostic interpretability, any operational deployment must account for persistence bias, regional variance, and calibrated uncertainty during climatic extremes.