Adaptive Early Wildfire Monitoring Based on Spatiotemporal Prediction and Himawari 8/9

Abstract

The rapid advancement of deep learning (DL) technology significantly enhances early forest fire detection methods. However, traditional approaches often rely on fixed thresholds and supervised learning techniques, which may fail to account for the complex spatiotemporal dynamics associated with forest fire events. To overcome this limitation, an adaptive DL model is proposed and specifically designed for early forest fire monitoring. This model integrates Stacking ConvLSTM to forecast mid-infrared brightness temperatures and employs a nonparametric dynamic thresholding method based on Otsu's algorithm for spatiotemporal anomaly detection, facilitating the identification of potential hotspots. By effectively capturing complex dependencies within spatiotemporal dimensions, this method improves detection accuracy. Furthermore, high-confidence early fire points are determined through dual-band analysis and land cover detection. Comparative experiments utilizing Himawari-8/9 satellite data reveal that the proposed method outperforms traditional techniques as well as the latest temporal methods, achieving an accuracy of 0.995, precision of 0.985, recall of 0.946, and an F1 score of 0.965. In addition, our method demonstrates an average fire detection delay of just 7 min and an average runtime of 71 s, underscoring its effectiveness in early forest fire detection. This approach serves as a robust tool for enhancing forest fire monitoring systems, offering significant implications for reducing response times and mitigating fire-related damages.