Automatic segmentation of satellite images based on a convolutional neural network

Abstract

The article deals with neural network methods for automatic segmentation of satellite images. The most popular Python libraries with neural networks for image segmentation based on Keras and TensorFlow are compared. In addition, a modification of the U-Net architecture is proposed, which optimizes the operation of the algorithm for the set task. To use the algorithm in satellite navigation systems, the definition of the algorithm is to be accurate without computational resources. The result is a change in the values of the loss function and metric during training.