Machine Learning based U - Net Algorithm for SAR Image Classification

Synthetic Aperture Radar (SAR) images are non-intuitive, and has a side-looking geometry which makes the interpretation of SAR images not very straight forward. Using automatic image annotation techniques, it is possible to annotate large quantity of images in less time with reliable outcome. Hence, the requirement of labelled SAR images is essential to improve the performance of Machine Learning (ML) algorithms. This effectively decreases the manual effort required. In this paper, we develop an unsupervised and semi - supervised algorithms for SAR image classification. A set of SAR images are downloaded from TerraSAR - X through open source repository and manually annotated using APEER annotation tool. The unsupervised learning is implemented using the traditional way of clustering and classification. It involves 48-D feature extraction using Gabor filters. The feature vector is labelled using clustering algorithms. K - Means, Gaussian Mixture Model (GMM), and Hierarchical Clustering methods are simulated. We propose a semi supervised learning that involves the implementation of a U-Net architecture for classification that involves clustering and segmentation. It is observed that the proposed methodology using U-Net is better than the conventional clustering. The U-Net model has an increase of approximately 64.224% and 56.43% in training and validation accuracy compared to the conventional clustering and classification.