Implementing a procedure to extract urban areas based on Multispectral Classification and Mathematical Morphology

We present an exploratory project aiming at progressing in studies concerning the following issue: when assessing the thematic classification accuracy, which should be the appropriate scale of the ground truth in relation to the resolution of satellite images used in the classification procedure? The work presented is the implementation stage of the procedures necessary to carry on this kind of studies with regard to the urban theme. The scheme developed includes: 1) the implementation of a classification procedure to extract urban areas based on multispectral classification and mathematical morphology, 2) a series of morphological transformations on the original vector map used as training and test areas, 3) the calculation of kappa index. The whole procedure has been satisfactory tested through a case-study. As to the problem stated above, the outcomes of the case study need to be further investigated. Research Questions This work is part of a larger study on the quality assessment of the urban object recognition in remote sensing image processing. General issue is : as urban objects are scale-dependent, which is the appropriate scale of the ground truth for which resolution of image processed in order to validate the result of image classification. To carry out the research, a procedure to extract built areas performed in our previous work (Bianchin, Pesaresi, 1993) has been implemented using GRASS and completed by a validation procedure of the result through Kappa index. The ground truth is provided by a vector map at scale 1:5 000, produced by the Regional Administration called CTRN (Carta Tecnica Regionale Numerica), which includes different layers (building, roads, etc). The building layer of the CTRN has been processed by using some morphological operators in order to obtain different resolution of the urban theme. The urban area of Treviso and its surroundings, in the Venetian region, has been selected as studycase. This is an area representative of the urban phenomena called "dispersive city". The procedure adopted The procedure adopted to extract urban areas from satellite image (Bianchin, Pesaresi, 1993) is explained by the schema below. It is based on two different processing chains for the two images: Landasat-TM and SPOT-Pan. Their results are merged to obtain an intersection image. The first chain is a two-theme (urban – no-urban) supervised classification of multispectral LandsatTM image. The second chain is based on mathematical morphological operators applied to the SPOT-Pan image to extract a ''structural' information related to the same theme. It starts from the monospectral SPOT-Pan image to generate the morphological gradient image PAN1. Building areas show a high value of the morphological gradient. PAN1 is re-sampled to be at the same resolution of TM so it may be intersected with TM. PAN2 is binarized by a threshold determined interactively on the basis of ground truth, to obtain a two-theme classification. PAN3 image and TM classified generate the intersection image IOUT. The basic idea of the procedure is to improve the accuracy of the TM classification trough the "structural" information extracted from the morphological processing of SPOT-Pan. Figure 1 – Scheme of the procedure The empirical case-study: Treviso In the area selected Treviso and its surroundings to experiment our research questions, image data available are: Landsat5 TM, 14 July 1994 (TM940714), SPOT-Pan , 17 August 1988 (063258), CTRN, 1:5 000, BIM-Piave, 1995, (sheets: 105 072, 105 083, 105 111, 105 112, 105 123, 105 124). The operating phases of the present job have been the following: · • pre-processing of the input data; • extraction of urban objects by a radiometric supervised classification of multispectral image Landsat-TM and analysis of its accuracy; • implementing the procedure of mathematical morphology to be applied to the SPOT-Pan image; • checking accuracy on the intersection image obtained. The research has been developed using the software MapInfo, Idrisi and Grass. We took advantage of the peculiar characteristics of each program in the different phases of job in order to improve efficiency and operating flexibility. The process for the extraction of urban areas has been developed entirely with Grass on Linux platform. The whole procedure has been automated by scripts containing iterative and conditional instructions. The stage of data pre-processing included the following operations: TM TM