U-Net Wildfire Classification in Sumatra Island Based on Sentinel-2 and Landsat-8 Satellite Images

Wildfire has become serious problem in Sumatra Island of Indonesia as its huge cost on damage of the surrounding environment. This work identifying wildfires via satellite imagery using U-Net, a deep learning model. Our method includes on using composite bands and calculating the delta normalized burned ratio from Landsat 8 and Sentinel 2 satellite images. Our steps of work has produced a pair of natural images with a delta normalized burned ratio (mask) and false-color images with a delta normalized burned ratio (masks). These two pairs of images are used to train the U-Net model. The results of the training using various pairs of images are compared with each other to achieve a reliable model for classifying wildfire areas. Our result shows that the best U-Net model was found with an accuracy of 0.94, a binary intersection over union of 0.78, a dice coefficient of 0.69, and a binary cross-entropy loss of 0.14. Our U-Net model can properly mark the pixel area of wildfires through satellite imagery.