Remote sensing image scene classification by transfer learning to augment the accuracy

Abstract

Efficiently executing image categorization with high spatial quality imagery from remote sensing can bring great benefits to scene classification. Effective feature representation is critical in the development of high-performance scene categorization techniques because sensor data processing is tough. Remote sensing as well as deep learning abilities have made it easier to extract spatiotemporal information for classification. Furthermore, other scientific disciplines, together with remote sensing, have made significant advances in image categorization by convolutional neural networks (CNNs), and transfer learning is being combined. Image categorization in this article was performed to enrich the accuracy of Scene classification using Transfer learning utilizing pre-trained Alex Net, and Visual Geometry Group (VGG) networks and compared with feature extraction methods. First, features were retrieved from the pre-trained network's second fully-connected layer and employed in SVM classification. Second, substituting the last layers of pre-trained networks with the notion of transfer learning was used to categorize new datasets. It is executed on the UCM Dataset as well as the SIRI-WHU Dataset. The proposed methodologies produced improved accuracy of 95% for UCM, 93% for SIRI-WHU Datasets.

Keywords

• Deep learning
• Feature extraction
• Image categorization
• Remote sensing
• Sensor data processing
• Transfer learning