SAR DATA ANALYSIS FOR LAND COVER CLASSIFICATION AND SOIL MOISTURE RETRIEVAL

Abstract

Satellite imaging is used to perform various vital functions for decision making especially for environmental modeling and Earth's surface monitoring due to its large geographical coverage and good penetration capability. Among various satellite sensor systems available these days, Synthetic Aperture Radar (SAR) has its own advantage because of its interaction and penetration capability with natural targets and day and night acquisition capability. SAR systems provide information about the amplitude as well as phase of the targets. This property of SAR systems lead to the concept of radar polarimetry. Polarimetric radar transmits waves with two orthogonal polarizations, viz., horizontal polarization (H) and vertical polarization (V), and receives the backscattered wave on either of polarization state. Thus, this radar system has channels: like-polarized channels, if polarization of the transmission and reception are the same, and cross –polarized channels, if polarization of the transmission and reception are orthogonal to one another. It results in four received channels, i.e., HH, HV, VV, and VH, where both the amplitude and relative phase are measured. So, a radar system can have different levels of polarization: Single polarized – HH or VV or HV or VH; Dual polarized – HH and HV, or VV and VH, or HH and VV; Quadrature Polarized– HH, VV, HV, and VH. The quadrature polarization radar is also called as fully polarimetric radar. In fully polarimetric radar systems, all the information in scattered field is retained by measuring the relative phase between channels along with the magnitude. Each polarized channel has its own significance. Cross polarized channel (HV) gives information about volume scattering. Double bounce scattering information is provided by HH channel whereas, VV polarized channel represents dominant surface scattering. In polarimetric SAR, the information regarding polarization, i.e., polarimetric information, shows much potential for characterization of land cover classes. The electromagnetic waves from the polarimetric SAR also have the significant capability of penetration into the vegetation which helps in acquiring information about electrical and physical properties of targets in terms of returned radar echoes, i.e., radar backscattering coefficient. The susceptibility of the radar backscattering coefficient of SAR data with the dielectric constant allows retrieving the soil moisture by using polarimetric SAR data. Soil moisture retrieval in case of vegetation covered areas is difficult because of the presence of multiple scattering effects existing in the canopy layer due to the presence of scattering contributions from vegetation components as well as from the underlying soil. ii Various works are available using SAR observables but still there is a need for model with lesser complexity for characterization of land cover classes and soil moisture retrieval. Therefore, in this thesis, an attempt has been made to develop a simplified adaptive land cover classification approach for PALSAR-2 data. Various change detection techniques are available to create change map, but to decide the threshold for change and no-change has to be much explored. Some techniques are simple and easy to interpret but cannot provide detailed change matrix. On the other hand, some techniques emphasize different information and reduce data redundancy but their results are difficult to interpret. Therefore, there is a need to develop such an approach which utilizes advantages of various existing change detection techniques, but minimizes the existing limitations. One of the major applications of SAR data is to estimate soil moisture especially soil moisture in vegetated cover areas because of its capability to penetrate into vegetation. Most of the techniques in this domain are based on the Water Cloud Model (WCM). Most of the existing works have estimated the soil moisture values by fusion of optical data information with microwave data. These methods require a priori information to a great extent as well as increase the complexity of system due to fusion process. Therefore, it is important to explore a method that utilizes single satellite data and requires minimum a priori information. Hence, a method has been developed which can estimate the soil moisture over vegetation covered areas by using information fusion of single satellite data and needs minimum a priori information. The main emphasis of this thesis is to develop such algorithms which can reduce computational complexity and should be adaptive in nature for application of land cover classification, Change Detection and Soil Moisture retrieval with fully polarimetric SAR data. The thesis consists of seven chapters. Chapter 1 presents the introduction to the basic concepts including motivation, major research gaps, and discussion about the study area with satellite data used. In Chapter 2, a brief literature review of related works, which is grouped into multiple sections, is presented. In Chapter 3, a critical analysis of commonly used polarimetric decomposition techniques namely, Pauli decomposition, Sinclair decomposition, Freeman and Durden decomposition, Van Zyl three-component decomposition, Yamaguchi four-component decomposition, Cloude decomposition, and H/A/Alpha decomposition for land cover classification has been carried out for PALSAR-2 data. Further, Maximum Likelihood supervised classification method is applied on iii decomposed images for classification purpose. A critical analysis of polarimetric decomposition techniques has shown that Van Zyl three-component decomposition method provides better classification results with overall accuracy of 93% and kappa coefficient value of 0.91 in comparison to all other decomposition methods. Yamaguchi four-component decomposition, and Freeman and Durden decomposition techniques also give the good result with overall accuracy 89.17% and 91.5%, respectively and kappa coefficient values of 0.87 and 0.89, respectively. For bare soil and short vegetation, the three decomposition techniques, Van Zyl three-component decomposition, Freeman and Durden decomposition, and Yamaguchi four-component decomposition show almost similar performance and performed significantly well. For the case of bare soil, the performance is found to be relatively better because surface scattering component can be easily discriminated with respect to the other types of scattering components. In case of tall vegetation, accuracy is low for Yamaguchi four-component decomposition technique and Freeman and Durden decomposition technique mainly due to the misclassification of the tall vegetation pixel into urban pixel and vegetation pixel. Cloude decomposition technique shows an overall accuracy of 75% and kappa value of 0.68. The decomposition performance of Pauli decomposition technique and Sinclair decomposition technique was found to be poor with overall accuracy of 68% and 65.84%, respectively, and kappa coefficient values of 0.59 and 0.57, respectively. H/A/Alpha decomposition method shows least classification accuracy of 48.1% and has kappa coefficient value of 0.34. Chapter 4 initially presents an approach for land cover classification using PDF (Probability Distribution Function) parameters. The results obtained from this approach are having overall accuracy of 79.35% and having kappa coefficient value of 0.734. It is observed that by using this algorithm, some pixels are remained unclassified pixels because of the presence of mixed class of pixels. Next, the chapter presents an adaptive land cover classification approach for fully polarimetric SAR data. In this chapter, various SAR observables, σhh, σhv, σvv, σhv/σvv, WPS, CPR, NDPI, RVI, are analyzed for segregating various land cover classes. We have selected optimal PolSAR observables for a particular class by using separability index criteria. In order to develop an adaptive approach, local information, i.e., information of whole images in the form of spatial statistics (mean and standard deviation) is utilized. For making the system adaptive, an empirical relationship between overall accuracy (OA) and spatial statistics of each selected polarimetric SAR observables is generated involving use of Genetic Algorithm (GA) for multi-objective optimization such that the value of OA is maximized. The algorithm has been applied on PALSAR-2 data. The iv classification results show good accuracy for each class. The advantage of the proposed approach is that it minimizes the use of apriori information. The overall accuracy of the proposed algorithm is 85.096% and kappa coefficient value is 0.81. The proposed approach has also been compared with Wishart Classification technique for which the accuracy was found to be 58.85%. Chapter 5 presents an approach for change detection that utilizes the change information present in different change detection maps. Various change detection techniques are available to create change detection maps, but the results are generally inconsistent. Therefore, change detection techniques from different categories namely, Algebra based, Transformation based, and Classification based are applied independently for change detection and thereafter change information from individual results of change detection map are used to detect change and no-change pixels. Firstly, Expectation Maximization (EM) algorithm is used for automatically selecting threshold for change detection map, after that information obtained from individual techniques is used to achieve final change detection map which gives confidence for change and "no-change" pixels. The algorithm is implemented on PALSAR-2 fully polarimetric SAR data. The proposed approach gives improved results as compared to involved base techniques showing the change and no-change pixels of the considered area. After fusing all three change detection results, we have observed that false alarm has been minimized and final change detection using proposed approach gives 15.08% change pixels and 84.92% no-change pixels. Chapter 6 deals with the development of soil moisture retrieval algorithm especially for vegetation covered soil moisture with fully polarimetric SAR data. WCM model has been modified using SPAN as parameter and soil moisture content were retrieved. The work discussed in this chapter identifies the problem of retrieving the soil moisture in vegetation covered areas by the use of fully polarimetric SAR observables. An algorithm has been proposed using HH, VV, and SPAN for soil moisture retrieval with PALSAR-1 data. The value of root mean square error (RMSE) is found to be 0.0309 and 0.0323 for retrieval of soil moisture when the algorithm is applied on test image and validating image, respectively. The proposed approach is also validated on PALSAR-2 data and it is observed from the results that Soil Moisture Content can be retrieved with RMSE value of 0.0324. Finally, Chapter 7 draws the conclusion and summarizes the contributions made in the thesis. The chapter also presents the future scope of the work carried out in the thesis.