Loss Function Selection in a Problem of Satellite Image Segmentation Using Convolutional Neural Network

Results of training a convolutional neural network for the satellite image segmentation are presented. Input images use four channels: Red, Green, Blue and Near-infrared. The convolutional neural network was trained to mark areas containing buildings and facilities. U-Net architecture was used for the task. For learning procedure supercomputer NVIDIA DGX-1 was used. The process of data augmentation is described. Results of training with different loss functions are compared. Network evaluation results for different types of residential areas are presented.