Histogram-based Segmentation of ALOS Polarimetric SAR Data

Mohammed Dabboor(1), Vassilia Karathanassi(2) and Alexander Braun(1)

(1) University of Calgary, 2500 University Drive NW, T2N 1N4 Calgary, Canada
(2) National Technical University of Athens, Heroon Polytechniou 9, Zographos, 15780 Athens, Greece


Full polarimetric data and their analysis methods provide detailed information on the backscattering behavior of objects. Backscattering behavior is affected by surface roughness and geometry, as well as humidity of the targets. Data segmentation methods, such as region growing and region split-merge are developed based on color and shape criteria which are the most suitable for electro-optical images. On the other hand, polarimetric SAR data and the images produced by the existing analysis methods, such as Pauli, Cloude-Pottier and Freeman-Durden, include information on the backscattering behavior of the different objects. Conventional segmentation approaches developed on electro-optical data do not take advantage of the additional information provided by the polarimetric analysis methods. The main problems encountered during conventional segmentation processing when full polarimetric SAR data are used, are the following: 1) the unions of the segment outlines do not correspond well to the edges of the different semantic objects, 2) it is necessary to use high scale values in order to segment the polarimetric images, as well as images produced by polarimetric analysis, 3) the produced segments, based on shape and “color” criteria, are not smooth.

This paper describes the development of a new segmentation method based on the analysis images calculated from the Pauli decomposition. Removing the speckle noise from the polarimetric SAR data by applying 7x7 Lee filter, the Pauli decomposition is applied in order to calculate the surface, double bounce and 45o tilted double bounce analysis images. In the first segmentation level, the data are segmented based on the dominant scattering behavior that appears in each pixel. The resulted segmentation contains three categories: the first category contains areas where surface scattering mechanism is dominant, the second one contains areas where the double bounce scattering mechanism is dominant, and the last one contains areas where 45o tilted double bounce scattering mechanism is dominant. In order to preserve the dominant scattering behavior, each category is further segmented. In the second segmentation level, three histograms are calculated for each category: the surface scattering, the double bounce scattering and the 45o tilted double bounce scattering histogram. 3D-histogram segmentation is performed using a non parametric histogram segmentation algorithm. Each histogram is divided into two regions and thus in the 3D-histogram space, each category is segmented into eight sub-regions. Following the same idea, third and fourth segmentation levels are further implemented. The resulted segments are evaluated using an electro-optical image of the study area.



  Higher level                 Last modified: 07.05.06