Texture Analysis for Correcting and Detecting Classification Structures in Urban Land Uses i

Transcription

1 Texture Analysis for Correcting and Detecting Classification Structures in Urban Land Uses i Metropolitan area case study Spain Bahaaeddin IZ Alhaddadª, Malcolm C. Burnsª and Josep Roca Claderaª ª Centre de Política de Sòl i Valoracions- CPSV, Universitat Polítècnica de Catalunya- UPC. Contact author: Bahaaeddin Alhaddad. C. Gran Capitán, 2-4, 3ª planta, Barcelona, SPAIN. bahaa.alhaddad@upc.edu, malcolm.burns@upc.edu and josep.roca@upc.edu Tel , Fax Abstract Texture analysis can be a good indicator of the presence of buildings and other objects and they are usually easier to detect than the often-complex multi-textured objects which cause them. Spot 5 images present complex scene of urban area. However, behind this complexity, within a focused area, buildings and industries tend to be aligned following some specific direction [1]. Elements of built-form, which together with land surface, physical infrastructure and communication networks comprise artificialised areas, tend to align in some dominant direction in a small area and possess geometric regularity. Therefore, their boundaries also align following these dominant directions in a small area in spite of the acquisition condition. Classification errors are caused by similar reflection (wave length) from different elements inside the satellite image, such as urban areas and irrigated land, affecting the separation between the builtform and non-built-form areas to define them in the classification process. What dose this mean? The amount of pixels in each category will play an important role in defining or increasing the accuracy of the final classification [2]. The resulting land activity classification of Spot 5 scenes covering the Metropolitan Areas (MA) of Madrid and Barcelona form the basis of this study. The result shows that the classification process confused similar parts of land cover. Moreover, unlike other classes in urban areas, the boundary can be successfully segmented by a conventional pixel-base classification method. The promising results from this analysis prove that an amount of pixels in each category boundary could be used as a potential cue for automated detection and for the correction of classification errors which had arisen in the process. This paper focuses on the development of a methodology based on the texture analysis of urban areas that may improve the urban investigation through remote sensing. This study can be divided into two fundamental steps: The first, to work over the initial classification results that showed an error between different elements such as urban areas and irrigated field areas that have a similar classification result. The idea is to apply texture analysis to separate the different elements by using a number of pixels in each category boundary. The second, Recovering isolated missing urban fabric data. In fact, it is hard to face this problem through the high resolution images for clearer illustration of all urban fabric areas but it is hard to get small elements that occupied small pixels areas to appear in the different classification process, in case these elements surroundings the different pixels that occupied with completely different elements. Texture analysis will play an important role in detecting this isolated data and reducing the error and improving the classification results [3]. I. INTRODUCTION The idea of applying texture analysis to urban areas originates from the consideration that an urban area can be defined on the basis of urban elements. The main advantage offered by this methodology concerns the contribution to estimate in the near future the seismic vulnerability in a region where data is not available or where it is difficult to collect [4]. The data sources are formed by Spot5 satellite images (False Colour (FC) images with 10m resolution, B&W image with 2.5m resolution). The FC image is used to classify urban areas and to create an urban land-use map, and to apply texture analysis on the urban class to obtain the missing information from the first classification result, then overlapped on the FC image in order to mask only missing classification parts. B&W image was used to improve the low resolution FC one to obtain Sparse Urban Fabric that could not be detected at low resolution classification result. This methodology provides answers to the following related points: Understanding the main classification problem for low resolution Spot satellite image.

2 Applying texture analysis on the urban class. Homogeneously urban texture result. Taking-off the missing information from the original sources. Applying new classification on the missing image, partly for correcting the result. Expanding the low resolution false colour image and using the B&W high resolution image to obtain the land cover of small categories (such as Sparse Urban Fabric). Overlay the missing information with an initial result for a final result. Finally, this paper will concentrate to illustrate further details of classification categories by detecting the urban areas and solving the main problems that could be faced through the high and low resolution satellite images that could give rise to an error or missing results for high accuracy urban land cover/use classification maps, as an application of remote sensing over the extensive urban metropolitan areas. II. OBSERVATION The land-cover classification process plays an important role in the study of urban areas. Spot 5 imagery provides sufficient details of information for urban mapping, but it also presents more complex scenes of urban areas. In such a scene, spectral information of building and industrial features, which is the focus of interest, becomes ill-defined. Conventional pixel-based image classification such as parallelepiped, maximum and minimum likelihood cannot produce a reliable land-use result in processing multi-spectral low resolution satellite imagery. The result obtained from the pixel-based image classification contained semi-correction for the classification results and the result contains 11 categories for the metropolitan area of Barcelona and Madrid. The first level of work carried out up to this point meant obtaining a general classification result and exploring which categories were contained in the image. However, some different categories were still mixed with the final result, such as the residential area with some irrigated fields, mineral extraction, and sites appearing as industrial areas; mixed data problems that had to be faced with the different kinds of satellite images through the classification process. Another principal problem that had to be faced was the low resolution images and the difficulty to find in the high resolution one the classification of small elements (isolated buildings) that occupied small pixels areas which were difficult to detect in the different classification processes, in the case of these elements being surrounded by different pixels containing completely different elements. The following sections will explain the different ways to improve the following results, to reflect greater details and higher accuracy in the final results. Figure 1. SPOT5 Madrid scene false colour, 10m resolution. The image illustrates a place of error, two different elements showing similar reflectance texture (Residential and irrigated field areas). SPOT Image Copyright 2004, CNES. Figure 2. A primarily result of supervised classification with 11 different categories. The image presents the error classification result. (Green circles illustrate the classification error and yellow circles present the correct classification for continuous residential development. Based on the above, an automatic texture analysis was proposed, which extracted urban fabric boundaries from low resolution images to correct (clean) the classification error. However, two methodologies were applied, related with texture analysis: the first for correction of the mixed results of land cover classification Fig. 1 & 2 ; and the second for recovering the missing isolated residential urban fabric areas Fig. 3 & 4. The methodologies were applied in order to speed up the location and the corresponding method to improve the low classification accuracy and the urban land use structure categories that could give an improved final result of the overall classification. Figure 3. SPOT5 Madrid scene false colours, 10m resolution. The image illustrates two different places: the city centre (yellow circle) and isolated residential area (red circle). SPOT Image Copyright 2004, CNES.

3 urban fabric has small boundaries that might appear as big closed points; the other land appears as polygons, so it is easy to remove take-off data of interest and remove the incorrect one [5]. Error in the classification result Texture Analysis to remove the information The correct interest area Figure 4. A first classification result shows the missing data of isolated residential urban fabric (red circles illustrate the classification error and yellow circles present the correct classification for residential urban fabric). (1) (2) (3) Change the Final categorie correct result Applying filters III. TEXTURE ANALYSIS FOR CORRECTION THE MIXED RESULTS OF LAND COVER CLASSIFICATION Many images contain regions characterised by variation in brightness rather than any unique value of brightness. Texture refers to the spatial variation of image tone as a function of scale. To be defined as a distinct textural area, the grey levels within the area must be more homogeneous as a unit than areas having a different texture. Buildings and industries possess different textures from irrigated land that sometimes had similar colours as indicated in the preceding figures. The complexity of urban areas also generates the complexity of texture analysis. There is considerably more textural density generated by buildings, industries or urban fabric areas, and less texture density generated by fields and large open space areas. In the first classification result the main error to be faced was that of the similar results for the irrigated fields and the urban fabric areas. The idea to separate between both different elements was the starting off point for the different occupation areas between the urban fabric areas, such as residential and industrial, and the open space, such as irrigated fields and dry lands. The following diagram illustrates the general concept of the methodology with the accompanying explanation. a) Diagram (1): The first classification result with irrigated land error reflects similar colour as the urban fabric areas. Two different steps were applied over this result: the first one, to apply the texture analysis to determine boundary elements; and the second, to change the urban fabric class to be of a similar colour as the irrigated field in the last case (it could be changed to be dry land if there were a common problem with other elements). b) Diagram (2): Use texture to apply occurrence based texture filters. Occurrence measures use the number of occurrences of each grey level within the processing window for the texture calculations. The remote sensing image processing tools such as ENVI illustrate different results between the urban fabric as small elements with complex texture and the open space areas such as the irrigated field ones. With low resolution images the (5) (6) c) Diagram (3): From the last step results a mask over Diagram (1) will be applied in order to remove the data of interest. Convolution median filters were applied while preserving edges larger than the kernel dimensions (these are good for removing salt and pepper noise or speckle). ENVI's median filter replaces each center pixel with the median value (not to be confused with the average) within the neighborhood specified by the filter size. d) Diagram (4): Applying the default, a 3 x 3 kernel, will remove most of the polygon areas and keep the close elements. e) Diagram (5): To overlap correct extraction data at origin the classification image must contain homogenous correct classes. In case of having such correct data to hand Diagram (4) the general error could be changed for the correct one. f) Diagram (6): The final will present the final overlapping and changing classification data. In all cases, the above process will improve the classification accuracy, but will not solve all of the problems. A. Methodology 1) Working with texture analysis Texture is a measure of the spatial variation in the grey levels in the image, as a function of scale. However, this process will apply just over the satellite images that have a resolution of less than 5m depending upon the above observation, that it is a way to use the small close polygons to take-off interest information and last concept will not be that useful in a high resolution images when the small elements such as building has enough size to illustrate outer edges as land size areas. Texture analysis will apply over classification result Fig. 5. (4)

4 2) Applying filters and works with look-up table Texture analysis presents black and white results with grey colour range in-between. The first result has a lot of noise with complex texture data especially over the urban fabric areas. It is important to work with B&W colour range to stretch it to present the interest areas, then applying a default 3 x 3 Kernel to remove the noise to clarify interest areas. 4) Change incorrect and overlay correct data In the case of residential areas having similar results as the irrigated fields (at the above case study area) and residential areas being clarify, from the last step the first result will be changed from the classification of the residential areas appearing as irrigated fields, with the correct extraction data being overlaid and the residential classification correctly assigned thereto Fig. 8 & 9. Figure 8. Homogenized incorrect irrigated field and correct residential areas. Figure 5. The following images present texture analysis to extract residential classification areas and remove the error classification of irrigated land areas that appears as a similar result. Applying filters over texture analysis results will be important to remove noises and clarify interest area. 3) Extraction Data of Interest From the last step, the texture result will use as a mask (white colour will present 1 DN and black colour will present 0 DN) over incorrect classification data to take-off the urban fabric areas and remove unneeded data Fig. 6 & 7. Figure 9. Overlay extraction residential areas. IV. TEXTURE ANALYSIS FOR RECOVERING ISOLATED MISSING RESIDENTIAL AREAS Figure 6. Texture analysis will present clear edge for big element areas. Spot5 10m resolution images defined some missing information at isolated residential urban fabric areas that did not appear as a result from the supervised classification process. Image processing tools such as ENVI or ER Mapper could not follow small pixel size at that resolution Fig. 3 & 4. In fact, it is hard to face this problem through the high resolution images. However, panchromatic Spot5 imagery for similar areas taken with a 2.5 m resolution was used to detect the missing data to improve the classification result. A. Methodology Figure 7. Applying texture mask will remove and keep data of interest. 1) Work with Panchromatic image Spot5 Panchromatic imagery with 2.5m resolution detects small details, and the clear edge areas mean a clear separation and identification of categories. One of the image processing benefits of geo-referenced pixels must occupy their position even though they have different resolutions. This process provided the opportunity to detect and observe data of interest. The following images Fig. 10 illustrate overlapping images

5 (panchromatic 2.5m and multispectral 10m) with more details of sparse (isolated) residential areas. Figure 10. SPOT5 Madrid scene stretched false colours 10m to 2.5m panchromatic imagery. SPOT Image Copyright 2004, CNES. 2) Working with texture analysis In the above correction methodolgy texture analysis was used to correct the classification result by applying the analysis over the classification image to improve the accuracy. In the following step texture analysis was used over the Panchromatic image with a similar concept as before (see 3.Texture Analysis for Correction the mixed results of land cover classification) where the sparse residential areas occupied a bigger area than before and appeared as closed polygon compared with other big elements such as agriculture fields. The following images Fig. 11 present two different areas: the first shows the sparse residential areas appearing as white points (closed polygons); the second shows mineral extraction areas appearing as open area (open polygon) after apply the texture analysis process. There is an interrelation between this step and the final one. Texture analysis was applied over the panchromatic image. The panchromatic image has clearer elements on the edges than the multispectral one, that help to extract new data sources for re-classification process. The texture analysis result from the last step was used as a mask over new data source image - the ovarlaping one - Fig. 10 to extract data of interest; then maximum likelihood classification was applied and reigons of interest (ROI) were used to choose the interesting samples. For the poor view and difficulty of applying ROI, band combination was applied to enhance the image improve the selection of data Fig. 12. Figure 12. Texture Analysis Mask to extract interest data from overlapping images; then band combination was used to enhance the view and aid in the selection of ROI. Maximum likelihood classification result from extraction data. 4) Overlay of the correct classification result ENVI tool offers an easy way for overlaying different data sources in case they have similar border size. However from the samples areas and the data sources that are used in this paper, the different border size allowed for the ER Mapper tool to be applied to this section of work by exporting only the data of interest (sparse residential area in this case) and applying the following formula to hide the background data (0 DN). (If I1 > 0 THEN I1 ELSE NULL) Where: I1: exported data (sparse residential area and background) The following images Fig. 13 show the selection area with the existing problem (undefined sparse residential area) and similar selection area after adding the extraction data. Figure 11. The small elements (red circle) that appeared in last classification result will have closed areas in the texture analysis for the extraction process. The big elements (yellow circle) will be removed from the data side. SPOT Image Copyright 2004, CNES. 3) Extraction and re-classification process Figure 13. Sparse residential classification areas existing with the problem solved.

6 The last process was undertaken automatically in order to save time. It could have a part of the error in some areas and less in others, but in all cases the accuracy of classification was improved and the formerly missing areas came into existence. V. CONCLUSION In order to take advantage of and make good use of remote sensing data, meaningful information needs to be extracted from the imagery. Interpretation and analysis of remote sensing imagery involves the identification and/or measurement of various targets in an image in order to extract such useful information. It has been shown that a technique that includes the user as an integral part of the process of detecting change can make the data capture more efficient. Automatic Texture analysis has been proposed, implemented and tested on Spot 5 satellite imagery. It was built to take advantage of texture existence in the scene and to convert texture data to a useful cue for correcting an error of the classification process. The test result showed high probability of texture analysis in the detection of land-cover with a high error of classification. Further testing in different sites can present more clearly the capability and limitation of texture analysis. It is recommended to integrate texture analysis, if applicable, into a fully developed automatic detection from low resolution imagery. However, in the first place, an attempt was made to try to understand the data sources and ascertain which information could be obtained. In addition, rather than ignoring the problems in the process, a concerted effort was made to solve the main problems with a high accuracy of result, such as the mixing and missing classification result. Moreover remote sensing tools were drawn upon in order to improve the details of the categories. In this paper, multivariate texture segmentation has been successfully used for improving the built-up classification data, concentrating on the residential urban fabric area as a case study that could be applied over other categories with data to obtain specific details. ACKNOWLEDGEMENTS The authors of this paper gratefully acknowledge the research funding provided by both the European Commission through the ERDF, by way of the INTERREG IIIB Programme, and the Spanish Ministry of Science and Technology (ref. BIA ). REFERENCES [1] Sohn, G. and Dowman, I., Extraction of buildings from high resolution satellite data. Automated Extraction of Man-Made Object from Aerial and Space Images (III). Balkema Publishers, [2] David, C. H. and Wang, X., Urban land cover classification from high resolution multi-spectral data over urban areas. Toronto, Canada, [3] T. T. Vu, M. Matsuoka, F. Yamazaki, Shadow analysis in assisting damage detection due to earthquakes from Quickbird imagery. Earthquake Disaster Mitigation Research Center (EDM), National Institute for Earth Science and Disaster Prevention (NIED), Wakinohama, Japan, [4] D. I. Morales, M. Moctezuma and F. Parmiggiani, Urban Edge Detection by Texture Analysis. National University of Mexico, Faculty of Engineering DEPFI-UNAM, Ciudad Universitaria. ISAO-CNR, Bologna, Italia, 2004 IEEE. [5] Anys, H., A. Bannari, D. C. He, and D. Morin, Texture analysis for the mapping of urban areas using airborne MEIS-II images. Proceedings of the First International Airborne Remote Sensing Conference and Exhibition, Strasbourg, France, Vol. 3, pp , 1994.