Abstract:

The preliminary results of the study on forest identification with Interforemetric Land use (ILU) image produced from ESA tandem data is presented in this paper. The contents relate to the selection of bands for building ILU image based on the aim of recognizing, the production of decision tree using Oblique Classifier 1 (OC1) and rectification of decision tree with our software, the evaluation of classification precision on pixel to pixel and the error analysis. ILU image is generally combined in the following way: the R (red) band is the coherence image, the G (green) band is the intensity image and the B (blue) band is the intensity difference image. The coherence image for one pair of SAR data is exclusive which means the R band is changeless, but the others might be different expressions. After synthetically comparing and analyzing for the histograms of different images and the classified images, it is fixed finally that the ILU image for recognizing forest and non forest is composed of the coherence image?, the minimum value image (G) and the standard deviation image (B). We developed a classification software based on decision tree since the decision tree produced from OC1 is not satisfied our need of classification. The processing steps are as following:(1) to collect the sample coordinates of various classes using Global Position System (GPS) in field work, (2) to get the sample data from ILU image according to their coordinates, (3) to remove the case of small probability and to create the decision tree using OC1, (4) to classify ILU image with the classification software based on decision tree and (5)to do post processing on classified image. The classification accuracy for forest is 77%, for non forest is 81% and for totals is 79%. The conclusions are as follows: After Interferometric processing, many kinds of ILU image based on different aim of application can be composed with two intensity images and one coherence image of a pair of ERS 1 and ERS 2 SAR image to be apart one day. The minimum value image and standard deviation image are better for distinguishing between forest area and water bodies than another intensity images and different images. There is no unclassified classes in the result dependent on the algorism of producing decision tree of OC1, the classes selected should cover data value range of image as many as possible and the sample data should be as “pure” as possible to reduce the error in the classified result.

Key words: ILU image, Interferometry, Forest mapping, Decision tree