A Hybrid U-Net and Attention-Based BiLSTM Framework for Wildfire Prediction Using Multi-Source Remote Sensing and Meteorological Sensor Data

Forest and grassland fires have become increasingly severe under climate change, posing significant threats to ecosystems and human safety. Accurate wildfire prediction using remote sensing data remains challenging due to complex spatiotemporal dynamics and heterogeneous data sources. To address this issue, this study proposes a hybrid deep learning framework integrating U-Net and an attention-enhanced bidirectional long short-term memory network (AUBLSTM) for spatiotemporal wildfire prediction using multi-source remote sensing and meteorological data. The U-Net is employed for spatial feature extraction, while AUBLSTM captures temporal dependencies and improves fire spread modeling with attention mechanisms. An encoder–decoder architecture is adopted to enhance multi-scale feature representation, and meteorological constraints are incorporated to improve physical consistency. Experimental results demonstrate that the proposed model outperforms baseline methods, including convolutional long short-term memory (ConvLSTM) and fully connected networks, achieving superior performance in terms of MSE, RRMSE, PSNR, SSIM, IoU, and F1-Score. The framework is efficient, scalable, and suitable for deployment in electronic monitoring and early warning systems, providing a practical solution for integrating multi-source data into wildfire surveillance applications.