Urban Change Detection From Spaceborne PolSAR Images With Radiometric Corrections

Since the last century, accelerated worldwide urbanization has greatly influenced the natural and ecological environments, society economy and the psychological lives of the citizens. Detection and monitoring of the land use and land cover types (LULC) in urban and around areas is arguable one of the most important applications of land observation using multi-temporal remote sensing data, i.e. optical or synthetic aperture radar (SAR) images. Satellite polarimeric SAR (PolSAR) can image the ground target in a2by2complex scattering matrix which can more accurately express the characteristic of each target objects, and make it possible for the target identification and information change detection of ground targets in different time.There are several problems in using PolSAR images for urban applications. First, when electromagnetic wave (EM) propagates through the ionosphere, faraday rotation (FR) effect causes significant impacts on long wavelength EM radiation, and result in the changes in scattering matrix. Second, the geometric distortions, i.e. layover, foreshorten and shadow, caused by the slant image geometry of SAR sensors in rough areas result in great radiometric changes of target in these areas. Finally, the limited polarimetric characteristics and the high correlation between which are also a problem for accurate change detection, and cause high demand of detection algorithm.For the achievement of high accuracy in urban change detection using PolSAR images, feasible radiometric calibration should be carried out and suitable detection algorithm are selected. Base on the principles of PolSAR signal propagation, polarimetric expression and polarimetric decomposition theory, the analysis of the effects of FR and rough topography are carried out, and the feasibility and reliability of polrimetric feature for change detection are discussed.First, for the impact of ionosphere FR effect, we analyze its intrinsic characters and the estimation algorithms of FR angle. Base on the polarimetric L-band phased array SAR (PALSAR) images acquired onboard advanced land observing satellite (ALOS), the FR angle is estimated using the circular polarimetric channels by the base transformation, and the radiometric calibration is done. This method can extract the angle in pixel-based level which is easily performed, and available results are achieved.Second, for the radiometric calibration in rough topographic areas of the ALOS PALSAR images, the principles of geometric distortions are discussed. The calibration methods of conventional types and that in rough topographic areas are described, and the proposed method based on facet area integration is highlighted. This method takes attention to the real interaction of PALSAR signals and the ground targets, and the calibration is based on the local illuminated area.Third, given the few polarimetric features and the high correlation between them, coherent/incoherent features and polarimetric decomposition features are extracted and divided into several groups. Two indices are used as the criterions for the selection of optimal feature group for target identification and change detection.Finally, different experiments are carried out. The supervised double-threshold segmentation based on canonical correlation analysis (CCA-EM) and the random forest (RF) algorithms are used for the extraction of change information. For samples selection and accuracy estimation, the Landsat TM images are used as ancillary data.Our research results of radiometric calibration concerning the FR effects and to geometric distortions is acceptable. A series of experiments are performed using multi-temporal ALOS PALSAR images and turn out that the effect of FR is quite small, which in general cases can be ignored. Calibration based on proposed area integration method shows obvious changes in hilly area, and the radiometric-terrain-calibrated (GTC) images are got. Change detection in Tai'an city using the proposed two methods based on the selected features get acceptable results in change information extraction, results using GTC images show little improvement corresponding to that using non-GTC images.