Scene classification in remote sensing images with deep learning

Deep learning is popular in remote sensing image scene classification as well as playing an important role in many fields. Remote sensing images provide automatic classification of the Earth's surface into various categories with data obtained from air and satellite platforms. In recent years, various approaches have been presented for scene classification. However, existing datasets have limitations in terms of class diversity and scale, which makes it difficult to develop new deep learning-based approaches. In this study, a pre-trained EfficientNet Version 2 Small (EfficientNetV2S) deep learning network was used for scene classification. By replacing the last layers of the model with a transfer learning method, a new deep learning model called EffiSceneNet was proposed and presented the best results in the literature. In addition to proposed method, the original EfficientNetV2S and EfficientNet Version 2 Medium (EfficientNetV2M) models were also run and the superior performance of the EffiSceneNet model was observed. Three models were combined with ensemble learning method with the proposed method. This method increased the overall accuracy rate by combining the individual performances of the models. Finally, in this study, Vision Transformer (ViT) model was also used for scene classification and showed performance. The study was carried out on large and diverse datasets and made significant contributions to the literature.