INDRA-Net: A Weakly Supervised Multiple Instance Learning Framework for Spatio-Temporally Interpretable and Extreme-Aware Crop Yield Forecasting in India

In agrarian economies like India, anticipating crop yield shocks before harvest is crucial for managing climate risks, stabilizing markets, and safeguarding food security. As extreme weather events become more frequent, policymakers need not only early warnings but also interpretable insights that explain where and why failures may occur. Yet, a key limitation remains: while remote sensing provides fine-scale information, yield data are usually available only at coarse administrative levels, and common averaging approaches erase the local variability that often drives yield losses. To bridge this gap, we introduce INDRA-Net (Interpretable Network for District Residual Aggregation), a weakly supervised Multiple Instance Learning (MIL) framework that directly predicts 38 years (1980-2017) of district-level yield residuals from high-resolution pixel-level time series. Unlike conventional methods that rely on naive spatial aggregation, the architecture employs a shared Temporal Fusion Transformer (TFT) backbone to independently encode the complex interactions between static drivers (e.g., soil properties, topography) and dynamic inputs (e.g., weather, vegetation indices) at the individual grid-cell level. These local embeddings are then synthesized via a learnable Gated Attention mechanism, which dynamically assigns higher weights to agriculturally relevant pixels while suppressing noise and non-crop signals. The framework is trained with a quantile regression objective to forecast yield anomalies, enabling explicit uncertainty estimates (P10–P90) essential for operational risk management. Extensive evaluation on wheat and maize yields across Uttar Pradesh, Punjab, Madhya Pradesh, and Bihar demonstrates that INDRA-Net reduces forecasting error (RMSE) by 12–14% compared to state-of-the-art machine learning baselines (Random Forest, XGBoost) and deep learning models (LSTM). By preserving pixel-level variability, the model captures localized extreme events—such as heatwaves or moisture stress- that are typically smoothed out by spatial aggregation. Crucially, the model’s three-dimensional interpretability aligns with crop physiology, correctly identifying maximum temperature during wheat grain-filling and precipitation anomalies during maize silking as the dominant temporal drivers, while isolating sub-district clusters responsible for yield failures. This enables the generation of granular yield anomaly maps without pixel-level labels, offering policymakers a scalable and operational tool for precision monitoring and targeted risk intervention.