EFFICIENT NEURAL NETWORK GPU CALCULATIONS IN THE TASK OF CLOUD SEGMENTATION IN OPTICAL SATELLITE IMAGES

The paper considers the problem of neural network segmentation of clouds in images generated by optical-electronic Earth observation satellites. To achieve high segmentation speed, the following measures are taken. First, the classic «U-Net» neural network architecture is optimized. Second, a library for executing convolutional neural networks (CNN) on GPU is developed. The paper describes the modified «Lanky U-Net» architecture, the format for internal storage of the neural network model as a computational graph, and the organization of computations within the graph. Parallelization of calculations on the GPU has been implemented. Experimental studies of the accuracy and speed of cloud segmentation in satellite images are presented.