Detection of forest disaster using high-resolution satellite images with CNN (convolutional neural networks)

Forests are important role in creating the atmosphere of the earth, and by absorbing and fixing huge amounts of carbon dioxide, the earth can maintain an appropriate atmospheric temperature. According to various domestic and international research results, forests perform functions such as reporting of life resources, prevention of desertification, and adjustment of micro climate. Especially, in Korea, 64% of the nation is made up of forests, among them, coniferous forests account for 43%, so the risk of forest fires is high. In addition, forest pest damage and hail damage have increased interest in forest disasters.

In this study, we conducted to detect the damage areas caused by forest disasters through high resolution satellite image data. There are many studies to extract the damages based on hyper-spectral aerial image, high resolution satellite image, vegetation index and factors affecting the forest environment. However, we tried to make it possible to apply more accurate and various forest disaster events by combining the deep learning technology. So we conducted research on large-scale fire disasters in Gangneung, Samcheok, and Sangj in May 2017, and hail-damaged areas in Hwasun, Jeonnam in July 2017. We collected image data on Sentinel-2 satellite and used U-net, one of the CNN(Convolutional Neural Network) model. Also, NDVI(Normalized Difference Vegetation Index), BAI(Burn Area Index), and FBI(Fire Burn Index) using the spectral characteristics of the normal forests and the damaged forest were used as the input data. We tried to verify the applicability to other forest disasters and to prove that CNN is a useful tool for segmentation and object recognition of satellite image data.