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

• Published in: International journal of applied earth observation and geoinformation Vol. 127; p. 103671
• Main Authors: Jin, Shuting, Wang, Tianxing, Huang, Huabing, Zheng, Xiaopo, Li, Tongwen, Guo, Zhou
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.03.2024; Elsevier
• Subjects: Fusion algorithm; Physical mechanism; U-Net; Wildfire detection; YOLO v5; Fusion algorithm; Wildfire detection; YOLO v5; Physical mechanism; U-Net
• ISSN: 1569-8432; 1872-826X
• DOI: 10.1016/j.jag.2024.103671

•Adaptive dynamic threshold is incorporated in the physical-based detections for generality.•The combination of physics mechanism and deep learning can remove system error from single algorithm effectively.•YOLOv5 framework is introduced firstly in heat source interference elimination of remote sensing fire detection.•The fire detection accuracy and applicability based on the fusion algorithm have been improved significantly. Currently, the spectra-based physical models and deep learning methods are frequently used to detect wildfires from remote sensing data. However, physical algorithms mainly rely on radiative transfer processes, which limit their effectiveness in detecting small and weak fires. On the other hand, deep learning methods usually lack mechanism constraints, thus generally resulting in false alarms of bright surfaces. It is promising to combine the advantages of them and correspondingly reduce the inherent error of a single algorithm. To this end, in this paper, both the local contextual and the global index method based on physical mechanisms are optimized, simultaneously, a new U-Net model is also establish to accurately detect fires. Moreover, YOLO v5 is incorporated for the first time to extract and remove the false alarms of objects with high exposure. Based on the above series of novel works, a self-adaptive fusing algorithm is finally proposed. Our results reveal that: (1) Short-wave infrared band of about 2.15 μm is crucial in fire detection for data with moderate-to-high resolutions. Taking Landsat 8 as an example, the band combinations of 7, 6, 2(SWIR + VI), 7, 6, 5(SWIR + NIR), and 7, 5, 3(SWIR + VI + NIR) show reasonable accuracy, with recall rate of greater than 81 %. The thermal infrared band can be used to assist in detecting the general location of the fire and serve as alternative choice in extreme cases. (2) The optimized physical algorithm can reduce false alarms and predict more accurate fire positions. (3) It is very effective to introduce the YOLO v5 framework to remove false alarms with high exposure in urban and suburban regions. (4) The proposed self-adaptive fusion algorithm integrates the advantages of various schemes, proving its better performance in terms of robustness, stability and generality compared to any single method. Even in extreme situations such as the Gobi Desert, thin cloud edges, and mountain shadow areas, the fusion algorithm still works well. The generality tests based on Sentinel-2A, WorldView-3, and SPOT-4 reveal the potential applicability of the newly proposed fusing algorithm, especially for data with fine spatial and spectral resolutions.