Object detection method based on instance segmentation of satellite image obtained in the conditions of cloud cover

Modern achievements in the space industry, combined with the continuous development of remote sensing technologies, form the basis for solving problems in various areas. Medium- and high-resolution satellite imagery often plays a key role in decision-making during crises in hard-to-reach areas. In the process of processing remote sensing data, a significant and still unresolved problem is the reconstruction of clouded images. This article analyses various approaches to cloud removal and data quality improvement. The traditional approaches considered have certain limitations associated with the loss of useful information. Particular attention is paid to deep learning methods, which are gaining popularity in solving cloud removal problems because they produce good results. The article discusses different DNN architectures (convolutional neural networks (CNN), conditional generative adversarial networks (cGAN)) and their modifications, identifies their advantages and disadvantages. A significant advantage of neural networks is their ability to adapt to various conditions and image types. The analysis of the disadvantages of fusing purely optical data led to the conclusion that the best solution to the problem of cloud removal from satellite images is to combine optical and radar data. As a result, the architecture of a model for removing clouds from optical satellite imagery using generative adversarial networks, combined with radar imagery, was developed. The theoretical hypotheses were confirmed by testing the model on the SEN12MS-CR dataset.