Kashmir-MungFormer: An Explainable Multimodal Transformer Framework for Green Gram Moisture-Stress and Yield Prediction under Temperate Conditions

Abstract

Green gram (Vigna radiata L.) is an important pulse crop whose productivity in the temperate Kashmir Valley is constrained by irregular rainfall, limited irrigation, cool-season variability, and episodic root-zone moisture deficits. Existing crop-yield models rarely integrate soil-water dynamics, weather, remote sensing, crop physiology, genotype, management, uncertainty, and decision support within a single leakage-aware framework. This study synthesizes the relevant evidence and proposes Kashmir-MungFormer, a physics-guided multimodal architecture for joint moisture-stress classification, phenological-state estimation, and yield prediction. The proposed system combines a temporal encoder for weather and soil-sensor sequences, an image encoder for satellite and unmanned aerial vehicle observations, genotype and management embeddings, reliability-aware cross-modal attention, and a differentiable soil-water-balance constraint. Multi-task learning, deep ensembles, conformal prediction, and explainability methods are specified for future multi-season field validation. An exploratory tabular benchmark was also conducted using 740 Moong (Green Gram) records from 26 Indian states with chronological model-fitting, calibration, and test partitions. Gradient Boosting produced the lowest test MAE (0.1625) and RMSE (0.2055) and the highest R² (0.3055), whereas Extra Trees produced the lowest MAPE (28.91%). Split-conformal intervals achieved 89.76% coverage at a nominal 90% level. The Jammu and Kashmir holdout yielded a negative R² (−0.1047), indicating substantial regional domain shift. Accordingly, the benchmark is reported as an exploratory baseline rather than a complete validation of the proposed multimodal architecture.