<![CDATA[Accurate forecasting of crop productivity is fundamental to contemporary food security planning, yet conventional predictive models frequently underperform when confronted with the multivariate, spatial, and temporal intricacies inherent in agronomic datasets. This study presents a robust deep learning framework leveraging a multivariate Long Short-Term Memory (LSTM) network to forecast yields of principal food crops. The model was developed using a panel dataset from 12 districts in Chhattisgarh and Madhya Pradesh, India (2010–2017), comprising area, production, and yield observations for multiple competing crops. Rigorous preprocessing protocols included the application of separate StandardScalers to mitigate matrix inversion issues, and the derivation of land-allocation features to capture spatial interactions among crops. A lightweight LSTM architecture stabilized by gradient clipping was employed to enhance convergence and prevent exploding gradients. Empirical results demonstrate that the multivariate LSTM notably outperforms simple baseline estimators by effectively modeling non-linear relationships and district-level yield heterogeneity, attaining an RMSE of 494.70 Kg/ha and an R² of 0.8031. These findings suggest that spatial anthropogenic indicators—particularly the allocation of land across commodities—serve as informative proxies for reliable yield prediction in contexts lacking comprehensive weather-sensor data]]>
Copyrights © 2026
