10-minute forest early wildfire detection: Fusing multi-type and multi-source information via recursive transformer

Abstract

Forest wildfire has great impacts on both nature and human society. While disrupts the ecosystems, wildfire leads to significant economic loss and poses a threat to local communities. To detect forest wildfire, remote sensing technology has become an essential and powerful tool. Compared with polar-orbiting satellite, the new generation of geostationary satellite provides higher temporal resolution and faster response capability. In this study, we utilize the near real-time data of Himawari-8/9 satellite, to achieve 10-min forest early wildfire detection. A recursive transformer model is proposed in this work. It fuses multi-type and multi-source information for Himawari-8/9 satellite. By leveraging the spectral, temporal and spatial features of fire pixels and considering land cover information, the proposed method reduces interference factors like cloud and terrain, resulting in minute-level and near real-time detection of forest wildfire. In 21 ground truth forest wildfire scenarios and MODIS-based cross-validation dataset, the proposed method achieves better results compared to the JAXA wildfire product, in terms of overall fire detection accuracy, early fire detection rate, omission rate, and real-time performance. Furthermore, the proposed framework effectively lowers the emergency response time for early forest wildfire detection, thereby reducing the loss caused by forest wildfire.