Multi-temporal remote sensing fire detection based on GBDT in Yunnan area

Abstract

How to detect forest fire on time is an important social problem. Remote sensing detection with large detection areas and high time resolution has become a popular method in detecting forest fire. Therefore, we propose a new method of detecting forest fire based on Himawari-8 geostationary satellite data. The detection algorithm used is the GBDT(Gradient Boost Decision Tree) machine learning combined with multi-temporal information referred to as MT-GBDT algorithm. Considering the characteristics of Yunnan Province, the algorithm of this paper uses contextual algorithms to filter potential fire points. Then, combined with multi-temporal information, we processed the potential fire point by using the diurnal temperature cycle (DTC) model. Finally, we used GBDT machine learning model to process the information obtained in the above steps. Result shows that the fire point information with high accuracy is obtained. The probability of false detection is 18%, and the probability of detection is 86%. Compared with the existing three methods, MT-GBDT algorithm successfully detects most forest fires, while other algorithms omit many forest fires, and the probability of detection is 40% to 50%.