Development and Training of a Neural Network Filter for Satellite Images Processing

Abstract

This paper is devoted to the study of the efficiency of using neural networks for filtering satellite images. The authors propose the use of convolutional noise suppressing autoencoders in order to minimize the filtering error variance. As part of the study, the architecture of the autoencoder was developed, optimal hyperparameters were selected and the resulting neural network model was trained. In addition, the paper compares the effectiveness of the proposed approach with traditional filtering algorithms such as Kalman filter and Wiener filter. Our models provide filtering efficiency gains of 3–4% at low noise levels (Signal-Noise-Ratio, SNR is 4 or more). The authors also investigated the effect of using data augmentations on improving the filtering quality. Experimental results showed that neural network models are able to outperform classical filters in terms of accuracy in processing real satellite images. Additionally, the paper studied the dependence of the filtering error variance on the number of training epochs of the neural network. The obtained results demonstrate that the developed neural network filter can be effectively applied for noise suppression on satellite images.