Change Detection in Polarimetric SAR Images Based on Improved Watershed Algorithm

Change detection using remotely sensed data has been used in many applications, such as the detection of dynamic changes in land cover and land use, the monitoring of forestland and agricultural land, the assessment of damage from natural disasters, and the study of urban environments. Despite the numerous studies devoted to multispectral and hyperspectral imagery, however, the use of optical imaging sensors is limited to weather conditions. Synthetic aperture radar (SAR) sensors can obtain daylight, cloud coverage, and weather-independent images. Their backscattered signals are also sensitive to the form, orientation, homogeneity, and surface conditions of a target. SAR imagery can therefore serve as a useful tool for detecting land cover and land use changes. The development of SAR techniques has given rise to Polarimetric SAR (PolSAR) systems, which measure four linear polarization channels (i.e., HH, HV, VH, VV) and the phase differences among them. SAR polarimetry with the functionality of identifying different scattering mechanisms can provide more significant information than single-channel imagery. From the perspective of image analysis unit, Change detection techniques are classified into two categories, namely, pixel- and object-based approaches. Pixel-based methods rely only on the information derived from individual pixels and do not consider the spatial relationships among these pixels. Whereas, the values of neighboring pixels in an image are strongly correlated, and the probability of change occurrence in adjacent regions is more than distinct points. The use of spatial features also effectively reduces the speckle effect in PolSAR images. Given these considerations, object-based approach has been widely used for PolSAR image analysis over recent years. This paper proposes a novel image segmentation algorithm for improving the accuracy of land cover change detection. This method consists of the following three steps: 1) segmenting two PolSAR images by new segmentation technique, namely region based improved watershed; 2) selecting the optimal differential of polarimetric features based on the Genetic Algorithm (GA) and Jefferies-Matusita (JM) distance criteria; and 3) The binary classification of image objects using the differential of mean pixel values of the corresponding image objects. Despite the development of various region-based segmentation methods, watershed segmentation is appropriate for the segmentation of high resolution images based on the many advantages of this morphological algorithm. These advantages include inherent simplicity, high speed implementation, the creation of separated regions in low contrast images, and the provision of closed connected regions. Common watershed segmentation approaches, such as distance transform and the gradient method, cause over-segmentation problem given the noise or local irregularities present in a gradient image. Unlike the direct application of the watershed algorithm, using a marker-controlled approach that involves the incorporation of additional knowledge into a segmentation procedure can limit the number of segmented regions. In this method, the flooding procedure begins from a previously defined set of markers. Markers, as connected components belonging to specific areas of an image, can be defined on the basis of a set of descriptors, such as gray level value, shape, location and texture. Compared with conventional watershed and multi-resolution segmentation methods, the improved watershed reduces the speckle effect in PolSAR images and avoids the over- segmentation problem. The results of proposed change detection method on Uninhabited Aerial Vehicle Synthetic Aperture (UAVSAR) full polarimetric images achieve the overall accuracy of 92.40% and the 0.85 kappa coefficient.