Integrated Hybrid AI–GIS Framework for Temperature-Driven Drought Early Warning and Agricultural Risk Mapping in Bangladesh

Drought has also become a potential threat of growing severity to agricultural sustainability and food security in climate-sensitive areas of South Asia. In Bangladesh, recurring and escalating episodes of drought in northwestern drought-prone areas, especially the Rajshahi district, have been caused by continuous heat-related stress, extended dry season, and changes in surface energy balance. In this research, an artificial intelligence-geospatial model is established to predict long-term temperature changes, measure spatial heat stress, and assess the agricultural risk posed by drought in the Rajshahi district. The ten hybrid forecasting models (LSTM, GRU, CNN, Informer, ARIMA, and ANFIS) that can combine both advanced signal decomposition methods (CEEMDAN, wavelet transform, variational mode decomposition, and STL) with deep learning and statistical learners were trained on a total of more than four decades (1980-2024) of monthly temperature records of the Bangladesh Meteorological Department (BMD). The results of the model performance were measured through various statistical values (MAE, RMSE, MSE, and R 2) and the temperature projections were made until 2034. At the same time, land surface temperature (LST) data of Landsat multi-decadal (1984-2024) was also run on Google Earth Engine to evaluate the spatiotemporal changes in warming. Findings indicate that warming has exhibited a statistically significant, constant trend, accompanied by a significant increase in high-temperature anomalies. The highest predictive accuracy models (RMSE < 0.12°C, R2> 0.99) were Wavelet-CNN-LSTM and CEEMDAN-SARIMA-CNN, while the remaining models, STL-ARIMA-BiLSTM and VMD-GRU-LSTM, also demonstrated good results (R2> 0.99). Transformer-based and shallow hybrid designs, on the other hand, had low reliability in long-term temperature prediction. Spatial LST analysis reveals a progressive growth and intensification of thermal hotspots, particularly in western and southern agricultural regions, indicating continued surface heating and a decline in thermal buffering ability. Clusters of high similarity between the predictions and actual patterns of surface warming contribute to the high risk of drought-prone agricultural areas and the growing stresses due to heat. In general, the suggested AI-GIS framework proposes a powerful, scalable system for drought early warning, agricultural risk mapping, and climate adaptation planning, delivering actionable information to support irrigation management and sustainable agriculture in Bangladesh and other climate-prone areas.