PolSAR Image Classification using Complex-Valued CNN and Squeeze-Excitation Network

Abstract

Terrain classification is one of the most crucial tasks when utilisation of polarimetric SAR images comes into the picture. This work explores the efficiency of various supervised deep learning algorithms that make use of Convolutional Neural Networks in land cover classification of PolSAR images. The goal of this work is to classify terrain into different ground covers such as urban, crops, forests, water, etc., from polarimetric SAR (PolSAR) images. State of the art classification approaches relish the advantage of deep learning techniques. However, conventional techniques such as convolutional neural networks (CNN), developed for optical images are not quite suitable for complexvalued PolSAR images. Hence, in this work, complexvalued CNN is employed to deal with complex values of PolSAR images. Further, the CNN focuses mainly on the spatial relationship within local receptive fields. However, the process entangles the channel correlation with spatial information. To address this issue, we use a squeezeexcitation network (SENet) along with complexvalued CNN to exploit the channel interdependencies. Thus, we utilize spatial as well as channel relationships in our work. This, in turn, helps in reducing the speckle noise in the images. Additionally, this work also tests the effectiveness of data augmentation techniques to increase the size of labeled training set of the three datasets used. This is done using various speckle noise suppression techniques. The experimental results on several datasets justify the importance of both spatial information as well as inter- channel correlation in classifying PolSAR images. The results after applying data augmentation techniques specific to the speckled nature of PolSAR images show improved performance of SENet architectures proposed in this work.