Cloud Removal in Satellite Images Using a Generative Adversarial Approach

Abstract

Automation of a wide range of operations in the agro-industrial sector at the current stage of technology development involves the application of advanced solutions in electronics, biotechnology, and information technologies. One of the directions of process optimization in agriculture is the introduction of information systems functioning on the basis of satellite imagery data analysis. To improve the efficiency of satellite data application, it is advisable to use machine learning and artificial intelligence methods. Satellite data allow monitoring a set of indicators describing chemical and physical characteristics, soil types, weather conditions, humidity. These indicators play an important role in the decision-making process of smart farming systems. The indicators are available in the form of satellite images, the quality of which depends on many factors. In order to apply such images in agricultural information systems, it is necessary to perform deep image analysis and cleaning. The approach aimed at applying a generative deep neural network to clean satellite images from clouds and shadows is proposed. The approach is based on training the neural network on synthesized data.