Wildfire Segmentation on Satellite Images using Deep Learning

Deep learning and convolutional neural network technologies are increasingly used in the problems of analysis, segmentation and recognition of objects in images. In this article a convolutional neural network for automated wildfire detection on high-resolution aerial photos is presented. Two databases of satellite RGB-images with different spatial resolution containing 1457 and 393 high-resolution images, respectively, were prepared for training and testing the neural network. Various techniques of data augmentation are used to enlarge training and test sets generated by data windowing. U-Net neural network with the ResNet34 as encoder was used in research. Neural network training was learning using the NVIDIA DGX-1 supercomputer. Adaptive moment estimation algorithm was used for optimization of training process. Special metrics, such as Sorensen-Dice coefficient, precision, recall, F1-score and IoU value allows to measure the quality of developed model. The developed algorithm can be successfully applied for early wildland fires detection in practical applications.