A study of algorithm to detect wildfire with edge of smoke plumes

Recent years, huge wildfires occur in many part of the world. And some researches have proceeded to improve wildfire detection with satellite imagery. Dozier (1981) developed the method that detects hotspot pixel by comparing the pixel with adjacent pixels. After that, Threshold method based on Dozier's approach and Contextual Method using relationship among neighbor pixels were appeared. But each of these algorithms needs more improvement in accuracy. In this study, we formulate a new algorithm with the edges of smoke plumes based on the rule of fire pixels match the origin of smoke plumes, and validate with the truth data. In this algorithm, MODIS band 1 (visible red) is extracted and smoke plumes are accented by histogram stretching. The edges of smoke plumes are extracted. Edge pixels that consist of fire smoke plumes are approximated by least squares method. Finally, the origins of the smoke plumes are determined and fire pixels are detected by the threshold approach. Our method, however, contain a problem that hotspot area shapes often a rectangle under the condition of not so high threshold temperature. In the results of this algorithm applied, it is found that it is easy to detect fire when clouds are not so thick and when smoke shape is visible clearly. On the other hand, false alarms along are detected along coast line and at the high refraction areas on a glacier, cirrocumulus clouds and so on. In addition, excessive detections increase in the low latitude because brightness temperature is raised by sunlight reflection. The wildfires in Alaska were detected well with our method. To validate this result, it is compared with the observational data and the common detection method. The Alaska Fire History Data (AFHD) is observed by Alaska Fire Service frequently, and the AFHD is offered as GIS data. On the other hand, MOD14 is one of the most famous and common methods to detect wildfire. It is calculated easily by MODIS data. Its accuracy rate to detect fire is high. So our results are compared with the observed AFHD and the MOD14 results. We compared the ratios of the accurate detections with both methods. Because our detected resolution is 250m and the resolution of MOD14 is 1000m, the pixels detected as wildfires with our method are almost four times of the ones with MOD14. Although they are quite different methods, the accurate ratios are similar: the correct detection with MOD14 is 87.2% and ours is 78.4%. Our method can detect some more wildfires that MOD14 cannot detect. In this result, our method must be able to supply the MOD14.