A self-adaptive wildfire detection algorithm by fusing physical and deep learning schemes

• DOI: 10.1016/j.jag.2024.103671
• Publication Date: 2024-01-27T10:17:22Z
• Authors: Shuting Jin; Tianxing Wang; Huabing Huang; Xiaopo Zheng; Tongwen Li; Zhou Guo
• Keywords: Fire Detection

Currently, the spectra-based physical models and deep learning methods are frequently used to detect wildfires from remote sensing data. However, algorithms mainly rely on radiative transfer processes, which limit their effectiveness in detecting small weak fires. On other hand, usually lack mechanism constraints, thus generally resulting false alarms of bright surfaces. It is promising combine advantages them correspondingly reduce inherent error a single algorithm. To this end, paper, both local contextual global index method based mechanisms optimized, simultaneously, new U-Net model also establish accurately Moreover, YOLO v5 incorporated for first time extract remove objects with high exposure. Based above series novel works, self-adaptive fusing algorithm finally proposed. Our results reveal that: (1) Short-wave infrared band about 2.15 μm crucial fire detection data moderate-to-high resolutions. Taking Landsat 8 as an example, combinations 7, 6, 2(SWIR + VI), 5(SWIR NIR), 5, 3(SWIR VI NIR) show reasonable accuracy, recall rate greater than 81 %. The thermal can be assist general location serve alternative choice extreme cases. (2) optimized predict more accurate positions. (3) very effective introduce framework exposure urban suburban regions. (4) proposed fusion integrates various schemes, proving its better performance terms robustness, stability generality compared any method. Even situations such Gobi Desert, thin cloud edges, mountain shadow areas, still works well. tests Sentinel-2A, WorldView-3, SPOT-4 potential applicability newly algorithm, especially fine spatial spectral