정지궤도 환경위성과 인공지능을 이용한 산불 연기 탐지

Abstract

Wildfires have been occurring for a long time due to natural and anthropogenic causes, but recently the frequency and intensity of wildfires have increased significantly. Wildfires are difficult to identify after they occur, and once they occur, they cause serious damage to the environment and society. Accordingly, various studies are being conducted with the goal of rapid and effective wildfire detection. However, research on large-scale pollution estimation using satellite remote sensing is still insufficient. This study aims to generate wildfire smoke detection products using GEMS(Geostationary Environment Monitoring Spectrometer) sensors. First, candidate pixels for wildfires were determined through the Swin transformer model, which has recently attracted attention in the image recognition field. In the next step, over-detection areas were suppressed through a denoising model created with a random forest, and a yellow dust classification model was built to exclude the possibility of false detection due to yellow dust among the finally detected pixels to generate wildfire smoke detection products. The evaluation accuracy of each built model was mIoU=0.858 and F1 score=0.920 for the Swin transformer, mIoU=0.908 and F1 score=0.926 for the denoising model, and mIoU=0.993 for the wildfire smoke and yellow dust classification model.