Classification Techniques of PolSAR Images

Abstract

Over the years, optical remote sensing technology has restricted the ability to capture images during harsh weather settings and at night time. However, Synthetic Aperture Radar (SAR) is independent of solar illumination and thus allows allweather continuous earth monitoring capability. A polarimetric synthetic aperture radar (PolSAR) is one type of SAR image which captures different attributes of the target by combining four different polarization states. Some PolSAR systems such as E-SAR, AIRSAR, F-SAR, etc., can capture abundant information of the target by employing multifrequency bands simultaneously providing rich information of the target. It makes SAR images suitable in wide range of Earth observation applications such as change detection, object detection, monitoring, classification, etc. This thesis addresses the classification problem of single frequency and multifrequency PolSAR images. PolSAR image classification is primarily a pixel based classification problem where our goal is to assign a label to each pixel of the image. Unlike optical images, PolSAR images are complex in nature which limits our ability of direct visual interpretation. Due to its active imaging nature, SAR images suffers from speckle noise which hinders the performance of pixel based classification. To address the classification problem, five contributions related to improving classification time and accuracy are discussed. We will begin with the introduction of Optimized Wishart Network (OWN) which is an improvement over the existing Wishart Network (WN) for the classification of single frequency PolSAR images. We propose methods to improve the classification time by reducing the computation overhead in WN and improve classification accuracy by proposing a better weight initialization method. Next, we propose the extended OWN (e-OWN) for classification of multifrequency PolSAR data. We show that the proposed method is able to combine different band information effectively and produces better classification accuracy. One of the big challenge for pixel based Pol-SAR image classification method is the presence of speckle noise in the image. To tackle that, we propose the superpixel driven OWN which uses both pixel and superpixel information to handle the noisy pixels. Finally, we present an stacked autoencoder based classification of multifrequency PolSAR images. All the proposed approaches are tested on variety of single frequency and multifrequency PolSAR datasets.