On classification of Sentinel-2 satellite images by a neural network ResNet-50

Abstract

Various combinations of neural network parameters and sets of input data for satellite image classification are considered in the article. The training set is completed with a NDVI (normalized difference vegetation index) and local binary patterns. Testing of classifiers created on a different number of epochs and samples is carried out. Values of the neural network hyperparameters are determined that allow a classification accuracy of 0.70 and an F-measure of 0.65 to be achieved. Separation into classes with similar spectral characteristics is shown to offer low classification quality at different parameters and input data sets. Additional information is required. For example, for forests to be divided into more detailed classes, one needs to employ classifiers that use images from different seasons and vegetation periods. In addition, the training set needs to be extended to take into account various natural zones, soils, etc.