Segmentation of landsat-8 images for burned area detection with deep learning

Abstract

Fires damage nature and living beings. Detection of this damage is important for future. In this study, it was aimed to determine burned areas. For this purpose, Landsat-8 images and U-Net model were used. Python language was preferred. Band combinations 7,5,4; 5,3,7; 5,4,3; 4,3,2; 4,3,2,5 and 2,3,4,5,6,7 have been tried. Train and test processes were carried out separately for each band combination. After the train and test processes were completed, a probability result consisting of values between 0-1 was obtained. Then, a threshold value was used. Thus, binary results consisting of 0 and 1 values were obtained. Three different values were preferred for the threshold: 0.1, 0.5 and 0.9. Thus, the effect of threshold value selection on the test results was examined. The prediction results were evaluated using the masks. For this, general accuracy, recall, precision, F1-score and Jaccard score metrics were used. Recall, precision, and F1score values were calculated for both burned areas and unburned areas. In addition, minimum, maximum, mean, and standard deviation values were calculated for each metric. When the results are examined, it is seen that the model gives better results when the threshold value is 0.1 and 0.5. Among the band combinations, it is seen that the 7,5,4 combination gave better results than the others. For this band combination, the highest mean accuracy is 0.9743 with the 0.5 threshold value. For this threshold mean recall, mean precision and mean F1-score for burned areas are 0.7203, 0.8411 and 0.7601, respectively. And Jaccard score is 0.6328.