Deep Learning-Based Remote Sensing Image Analysis for Wildfire Risk Evaluation and Monitoring

Wildfires have significant ecological, social, and economic impacts, release large amounts of pollutants, and pose a threat to human health. Although deep learning models outperform traditional methods in predicting wildfires, their accuracy drops to about 90% when using remotely sensed data. To effectively monitor and predict fires, this project aims to develop deep learning models capable of processing multivariate remotely sensed global data in real time. This project innovatively uses SimpleGAN, SparseGAN, and CGAN combined with sliding windows for data augmentation. Among these, CGAN demonstrates superior performance. Additionally, for the prediction classification task, U-Net, ConvLSTM, and Attention ConvLSTM are explored, achieving accuracies of 94.53%, 95.85%, and 93.40%, respectively, with ConvLSTM showing the best performance. The study focuses on a region in the Republic of the Congo, where predictions were made and compared with future data. The results showed significant overlap, highlighting the model's effectiveness. Furthermore, the functionality developed in this study can be extended to medical imaging and other applications involving high-precision remote-sensing images.