Yield Estimation of Paddy and Maize Crops in Mahabubabad District Using Sentinel-2 Data with a Hybrid Convolutional Neural Network-Lstm Model

Abstract

Estimating crop yield during the early growing season is important for agricultural planning and food security, particularly in semi-arid regions where climate variability plays an important role in influencing crop production. In this study, a novel deep learning framework that combines multi-temporal Sentinel-2 imagery is developed for estimating the paddy and maize yields during the early growth stages for Mahabubabad district, Telangana. The methodological framework involves Convolutional Neural Networks (CNN) to extract spatial features and Long Short-Term Memory (LSTM) networks to analyze temporal patterns, combined with an attention mechanism to identify key phenological stages. The model is trained with 874 field polygons of field data, spectral and phenological information from multi-temporal Sentinel-2 during 2023-24. Hybrid CNN-LSTM architecture results demonstrate that the yield prediction achieved an R² of 0.87 for paddy and 0.84 for maize at the maturity phase, which is higher compared to traditional machine learning approaches. Temporal analysis indicates that prediction accuracy varies considerably during different phenological stages, with the best performance at 45–60 days before harvest. The developed model exhibited good early-season prediction ability with highest accuracy during the reproductive phase of paddy and maize crops. Overall, the proposed methodology offers a scalable framework for early-season yield estimation in semi-arid climates.