A Polarimetric SAR Classification and comparison test against aerial photography images in Glen Affric radar project

In this paper our recent investigation and experimental results of the polarimetric SAR classification of the data for the Glen Affric radar project is presented. The polarimetric classification is one of the most important applications of the radar polarimetry in remote sensing. The unsupervised and supervised polarimetric classification of the Glen Affric data in respect to the classified Arial photography images of the same site using the Wishart classification technique is presented. The focus of the classification is on the diversity of the natural/semi natural landscape with a particular view on the wetlands/moor lands and lochs. The data for this project is acquired under the UK NERC / BNSC sponsored SHAC campaign. Introduction As part of the Biodiversity action plan and to comply with recent European legislation requirements on the environment, the effective monitoring, of regenerating and preservation of the natural pine wood forests has gained major significance. Classification of semi natural forested areas in an undulating landscape is a challenging remote sensing task. The knowledge of vegetation type, landscape cover and the classification of the morphological characteristics of the land cover structure is not only an important environmental issue but also has a great economic value. In this paper two polarimetric classifications of the test site radar data based on the target parameters decomposition is presented. We also show that the polarisation orientation angle ‘β’ can introduce a factor for further classifying the low entropy/ low alpha structures. The Wishart classification is used throughout to iterate and enhance the similarities of the backscatters with respect to their coherence and structure. The radar classified results are compared and tested against an optical classified image of the same area. The comparison shows the rapid convergence requiring up to 5 iterations. Experimental data set description The SHAC campaign was carried out in April 2000 under the UK/NERC/BNSC sponsored project. The radar data is fully polarimetric L Band. The image chosen for the comparison of optical and SAR classifications, was a Landsat TM scene (206/20) acquired on the 5 May 2000. This image was appropriate to the study as it provided spectral vegetation information of a comparable date to the SAR data. The image was geometrically corrected to a geographic (Latitude/Longitude) map model, through the selection of ground control points. A second order polynomial transformation was used to model the relationship between image coordinates and the geographic map co-ordinates. Control points with large residuals were iteratively deleted until the final residuals ranged from 0 to + or – 0.75 pixels. A nearest neighbour interpolation was used to resample the original image to the desired output pixel locations in decimal degree units. These and all other optical image processing procedures used in this study were performed using ERDAS Image Processing Software, (ERDAS, 1999). The Landsat TM image used in this study has seven spectral bands. The Ground reference data is the LCS88. It is used as a hybrid of ecological and cultural classes, based on strict boundaries with no regard for spectral properties and shows considerable heterogeneity. Meanings to the optical spectral classes were assigned afterwards using visual human interaction with ground reference information (LCS88). The optical classification produced 30 separate spectral cover features, which were then associated into greater spectral groups, such as water, woodland (plantation and semi-natural), undifferentiated heather and peaty Mooreland, rough grazing, and bare ground. Experimental data site description The Glen Affric site is centered at 4.90 west to 57.28 north. The radar image ground resolution is 2.18m by 1.8m in range and azimuth planes respectively. The total ground image size spans 10.4Km by 3.2Km. Fig. 1. Map of the Spectral classification of Glen Affric Background review The polarimetric classification of radar images based on target decomposition parameters ‘Alpha’, ‘Entropy’ and the second order statistics of the radar coherence parameters were suggested by S.Cloude & E.Pottier. [1] This classification method was combined with the Maximum likelihood Wishart classifier. [2]. The alpha entropy zoning classification provided the training set of data for the complex Wishart classifier. The complex Wishart classifier is based on two operations. The first is the mean coherence and the second is based on the structural similarities of the class members. In each iteration the image pixels are trained against the mean class coherence. The minimum distance between centres of a pixel and the class cluster is given by: ( ) m m m T T T Trace d det . 1 + = − for any dm < di , 0 45 is mainly associates with smaller lochs with shallow water. Fig. 6 shows the map of the ‘β’ parameters used at the initial training set of the unsupervised classification The initial Alpha/Entropy/Anisotropy classification of the Glen Affric radar data formd 18 classes and had a convergence rate of less than 5 iterations. The Anisotropy threshold limit was set to its mean value of 0.325. Z1 Z2 Z3 Z4 Z5 Z6 Z7 Z8