Scene classification of Google Earth Images with different deep learning models

Abstract

Deep Convolutional Neural Networks are widely used for the automatic labeling of satellite imagery. Four different CNN models were trained using the AID dataset, which is an openly shared dataset, and the performances of the trained models were tested using the general accuracy metric. Test performances of ESA models were compared with similar studies in the literature on the same data set. At the same time, the model transferability of the trained models with the domain-shift application was tested with the help of different images obtained from the Google Earth platform. In this way, it is aimed to expand the application of automatic labeling of land use and land cover classes with deep CNN models in different data. The test results also support that the use of semantic scene classification algorithms is becoming more and more promising with the developing technology and opportunities.