Improving accuracy of land cover classification through ensemble classification of lightweight deep learning models on EuroSAT dataset

Abstract

The object of this research is the classification of land surface types in multispectral images from the EuroSAT dataset. The subject of the study is high-precision algorithms for classifying images with thirteen channels, characterized by a small amount of computation. The use of such algorithms for processing images by onboard computers of the small satellites of the Sentinel-2 constellation, launched as part of the Copernicus program of the European Space Agency for Environment and Security, will improve classification accuracy and reduce the required power consumption, memory capacity, and the load on communication channels. Given the exponential growth in the required number of parameters of known neural networks to improve classification quality, an alternative approach is proposed in this paper. It involves the use of ensemble classifiers with stacking. In the first stage, it is proposed to use the most lightweight versions of neural networks with a relatively high level of classification quality. In the second stage, a multilayer perceptron is used. In this paper, lightweight versions of well-known neural networks are constructed: ResNet, EfficientNet, Compact SS-GhostNet, FastViT, and CCT. Using them within a stacked ensemble classifier yielded a classification quality of ninety-nine point fifty-nine percent on the EuroSAT dataset. This is zero point four percent higher than the best classification quality achieved by the CALCNet neural network. Moreover, the number of parameters in the proposed ensemble classifier is five times smaller than that of CALCNet. Versions of all programs are available on GitHub.