Abstract: Focused on studying wildfire and driving forces across various ecosystems on a large scale, this paper presents a novel method utilizing the google earth engine (GEE) platform. Firstly, the fire information for resource management system database, Sentinel-2 images and driving factor information in four different Chinese ecosystems were accessed on online via GEE platform. Then, the differential normalized burn ratio was extracted from Sentinel-2 images to sieve fire spots. Three machine learning algorithms, namely random forest, support vector machine and augmented regression tree, were used to classify fire locations. Furthermore, we determined the optimal-performing classification algorithms for each ecosystem and assessed variable importance. The results showed that random forest performed best with accuracy exceeding 92% among the three machine methods and the fire drivers varied significantly among four different ecosystems. In Changzhi City of Shanxi Province, and the Great Xing′an Mountains of Inner Mongolia, population distribution and maximum temperature were identified as the most influential drivers, respectively. While for Liangshan Yi Autonomous Prefecture in Sichuan Province and Ganzhou City in Jiangxi Province, the palmer drought index and soil moisture emerged as the primary drivers. This study demonstrates the efficacy of the proposed GEE-based method in studying wildfire and driving forces across different ecosystems in large scale regions.

Key words: wildfire, driving force, random forest, support vector machines, enhanced return tree, google earth engine (GEE)