Image segmentation for burned area detection from satellite imagery using the U-Net deep learning model

Fires threaten life all over the world and damage millions of hectares of area every year. Remote sensing provides advantages for damage detection in terms of time and cost. By using satellite imagery, burned areas can be detected without the need to visit the area. Since factors such as image band configuration, optimisation algorithms, and thresholds affect the results, this study aims to observe their impact on burned area detection. Thus, by using Landsat-8 images and U-Net architecture through the Python programming language, various combinations were created and different thresholds were used. According to the results, the combination of 7, 5, 4 bands and the AdaMax algorithm were selected for the final model, and the results were improved by data augmentation. Consequently, accuracy obtained in the final model was 97.76%, which was the highest for a threshold of 0.5. The F1 score obtained for the same threshold was 79.38%.