Automated classification of remote sensing images using multileveled MobileNetV2 and DWT techniques

Automated classification of remote sensing images is one of the complex issues in robotics and machine learning fields. Many models have been proposed for remote sensing image classification (RSIC) to obtain high classification performance. The objective of this study are twofold. First, to create a new space object image collection as such a dataset is not currently available. Second, propose a novel RSIC model to yield highest classification performance using our newly created dataset. Our presented automated classification model consists of multilevel deep feature generation, iterative feature selection, and classification steps. The features are extracted from the images using pre-trained MobileNetV2 and discrete wavelet transform (DWT) methods. The combination of DWT and MobileNetV2 generates large number of features. Then, iterative neighborhood component analysis (INCA) is used to select the best features. Finally, selected features are fed to support vector machine (SVM) for automated classification. The presented model is validated using two RSIC datasets: UC-Merced, and newly created space object images (publicly available at: http://web.firat.edu.tr/turkertuncer/space_object.rar). The developed model has obtained an accuracy of 98.10% and 95.95% using UC-Merced, and newly generated space object image datasets, respectively with 10-fold cross-validation strategy. It can be concluded from the results that, the presented RSIC model is accurate and ready for real-world applications.

Automated summary

The study created a new dataset of space object images and proposed a novel RSIC model, which achieved the highest classification performance using this dataset. The model utilized pre-trained MobileNetV2 and DWT methods to extract features, selected the best features through the INCA algorithm, and classified the images using SVM.