An Enhanced Hybrid LSTM–Linear Regression Framework for 90-Day Rainfall Forecasting in Rainfed Agricultural Regions

This study develops a 90-day rainfall forecasting model for rain-fed agricultural systems in areas lacking ground-based weather sensors. To address this limitation, open data from the NASA POWER satellite (2014–2024) covering Nadi Subdistrict, Mueang District, Surin Province, Thailand, is utilized. The proposed hybrid framework combines Long Short-Term Memory (LSTM) networks with Linear Regression (LR) to join the ability to learn nonlinear temporal dynamics with the potential to model long-term trends, complemented by lag and rolling windows techniques. The results indicate that the lag–rolling–augmented LSTM–LR model achieved the lowest Root Mean Squared Error (RMSE) of 0.0413 and Mean Absolute Error (MAE) of 0.0230 among all 10 models. The Friedman test confirmed the significant difference (x² = 364.33, p < 0.001), and the Nemenyi test showed that the proposed model significantly outperformed the traditional BiLSTM, Convolutional Neural Network (CNN), CNN (with lag and rolling features), and LSTM+LR (p < 0.0001). Furthermore, the model maintained high accuracy during both heavy and drought conditions. The novelty of this study lies in the integration of temporal hybrid models with open satellite data to create a low-cost and statistically validated tool for climate risk management and decision support approaches for smallholder agriculture.