Investigation of the Impact of Primary Data Processing on the Results of Neural Network Training for Satellite Imagery Recognition

Abstract

AbstractResearch and recognition of high-resolution multi-spectral satellite images is a complex and vital task in modern science. Of particular difficulty is the recognition of satellite images without additional channels, based only on RGB, in addition to very homogeneous classes with similar primary features. This article discusses the creation and configuration of a neural network architecture based on ConvNet with subsequent training. High-resolution satellite images from the Landsat 8-9 OLI/TIRS C2 L2 without additional channels to create datasets necessary for the neural network to complete the task. The paper presents four experiments on ranking the input data of the neural network to identify their influence on the final result of recognition, regardless of the settings and architecture of the neural network itself. Pre-processing of the input data was based on annotating the images of each class and then creating masks for them, namely for the classes: water, trees, and fields. In experiments with prepared validation input data, the increase in class recognition was up to 54.44%. Conclusions have about each experiment and the influence of input data on the result of satellite image processing. KeywordsNeural networkImage recognitionData processingSatellite imageryArtificial intelligence