Classification of Synthetic Aperture Radar-Ground Range Detected Image Using Advanced Convolution Neural Networks

Abstract

Synthetic aperture radar image data with advanced algorithms have opened up many layers of opportunities and applications in the field of SAR image processing such as individual object’s classification etc. This paper presents pixel-based convolution neural network (CNN) models that are useful for the high-resolution synthetic aperture radar (SAR) images of Ground Range Detected (GRD) data or image of city or town, agriculture and surrounding areas which contain other features for a complete, quick as well as the exact categorization. The land area which has been investigated in this paper is labeled into seven categories: viz., forest, water bodies, settlements, agriculture, vegetation, sand, and open area. The accuracy of the classification was compared to the other regions of land that have similar categories. An entire classification outcome on opted sections of the imagery would be showing that, pixels based convolutional neural network models of AlexNet, GoogLeNet, and ResNet50 are viable to solve both task of object identification and classification. The performance is evaluated in terms of training, testing accuracy, and kappa coefficient or confusion matrices for the GRD SAR image. The overall accuracies of greater than90% are achieved with the CNN techniques.