Machine learning apporaches to wildfire detection using geostationary satellite data

Abstract

More than 60% of the areas in South Korea consists of forests with rugged terrains. Thus, when wildfires occur, they can rapidly spread out. In addition, most wildfires are caused by anthropogenic factors, which are unpredictable. Early detection of wildfires can reduce the related damage. This study proposed a novel ensemble approach for detecting wildfires using Himawari-8 satellite data and machine learning in South Korea. Himawari-8 Advanced Himawari Imager (AHI) is the geostationary satellite sensor operated by the Japan Meteorological Agency. AHI collects data at 16 bands from visible to infrared at 500 m – 2 km resolution covering from East Asia to Australia every 10 minutes. In-situ wildfire data provided by the Korea Forest Service were used as reference data. The proposed approach combines a threshold-based algorithm and random forest machine learning. The threshold-based algorithm first detects potential wildfire areas and then random forest is adopted to remove false alarms from the identified potential wildfire areas. The 3.85 band widely used in wildfire detection was used in the first step in the threshold-based algorithm. After the first thresholding step, three conditions based on the characteristics of fire and non-fire pixels were used as a second thresholding step. The results from the threshold-based algorithm are fed into random forest machine learning. At first, all band reflectance, brightness temperature, ratio and difference values were used as input variables to random forest. Based on the relative variable importance identified by random forest, a total of 26 variables were finally determined as input variables in the random forest model for wildfire detection. Since time series data with a 10 min interval were used, not only detecting wildfires, but also monitoring them was conducted. In particular, how early the proposed approach identified wildfires was examined. Results showed that the proposed algorithm detected wildfires much better than the existing ones, especially for small-scale wildfires. Among total 65 reference cases, 50 cases are detected within 22 minutes except one case. In addition, false alarm rates found in the existing algorithms (e.g., MODIS hot spots), were greatly reduced when using the proposed approach.