INFORMATION CONTENT ANALYSIS OF KVR-1000 ORTHO-IMAGE BASED ON THE AVAILABLE TOPOGRAPHIC MAPS IN THE GIS ENVIRONMENT

Transcription

1 EARSEL Workshop on Remote Sensing for Developing Countries, Cairo, INFORMATION CONTENT ANALYSIS OF KVR-1000 ORTHO-IMAGE BASED ON THE AVAILABLE TOPOGRAPHIC MAPS IN THE GIS ENVIRONMENT H. Sahin, G. Buyuksalih, H. Akcin, H. Topan, S. Karakis, A. M. Marangoz Karaelmas University, Faculty of Engineering, Dept. of Geodesy and Photogrammetry, Zonguldak, Turkey; sahin@karaelmas.edu.tr ABSTRACT Recent developments in remote sensing enable us to use the high resolution satellite images for different purposes. Main application field for remotely sensed data is the large-scale topographic mapping and map revisions for which it is necessary to detect and extract the features from such imagery. In this study, automatic and manual feature extractions have been realized on the KVR-1000 orthoimage of Zonguldak testfield in GIS environment with the comparison of available map materials. As a result, while only buildings, roads, water structures and forest classes can be efficiently extracted in an automatic fashion, total information for 1:5000 1:25000 scale maps can be compiled with the help of manual extraction. Additionally, 71% and 78% of the buildings can respectively be extracted automatically and manually from the orthoimage. INTRODUCTION The archive of various Russian space cameras includes images acquired in more than 20 years. In particular, the image data has been taken by TK-350 camera since 1981 and KVR-1000 camera since For a long time, images from these cameras were used only by Russian organizations for mapping purposes. However, these images became available for a wide circle of users since 1991 with the unexpected decision of Russian government. This caused the easing of restriction from US government on American space imagery to 10m. Thus, licences were issued to several American-based companies to develop commercial satellites producing imagery to the 1m ground pixel level (Petrie, 1999; Li, 2000). The right of commercial distribution of KVR-1000 images is reserved by the company called Sovinformsputnik (SIS), which is the first Russian company of such type (see and massive global coverage exists from the system, with imagery from , 1998 and Such a long archive could provide valuable source for change detection applications which allows monitoring the past and present situations over the interest area. However, KVR image data are potentially used for updating and creating all planimetric details for the orthophoto plans with a scale 1:10000 and smaller (Chekalin and Fomtchenko, 2000; Lavrov, 1996 and 2000). According to Konency et al. (1982), for the detection and the identification of topographic features for vector line mapping at a scale of 1:50000, a ground pixel size of 3m for monoscopic imaging or 6m for stereoscopic viewing is necessary for a single building. For the interpretation of areal features such as forest, cultivated areas or vegetation or large linear features such as streets or rivers, larger pixel sizes can be accepted. In Topan, et al., 2004, a rule of thumb for the relation of the pixel size and the map scale is given as 0.05 up to 0.1mm pixel size in the map scale that means for a map 1 : a pixel size on the ground of 2.5m up to 5m is required. For a good perception of the detail present in an orthoimage, Doyle (1984) recommends a maximum scale which is achieved by multiplying the ground pixel size by the factor of 10,000. But the pixel size is not the only criteria for the quality of the images; also the contrast (modulation transfer function) is important like the spectral range and colour information. This may also dependent upon the situation of the atmosphere and the sun elevation. In addition

2 EARSEL Workshop on Remote Sensing for Developing Countries, Cairo, the area may be different we may have some wide roads and large buildings like for example in the USA or in Saudi Arabia or we may have small and bending roads without pavement. Also the specified information contents for the maps may be different for example in Switzerland we do have a lot of details in the maps, in the USA the maps are quite more general. So there is still a range within the relation map scale to pixel size on the ground. In this paper, the authors first report on the results of the geometric accuracy testing of KVR-1000 digital orthoimage. For this purpose, Tran3d program from Hannover University (including shift only, similarity and affine transformations), was implemented using the image and ground coordinates of GPCs distributed uniformly over the orthoimage of Zonguldak testfield. Accuracy results attainable by each transformation are reported and discussed individually with the detailed analysis to show the effectiveness of KVR-1000 orthoimage for large-scale mapping applications. In this case, automatic and manual feature extraction capabilities for different object in a view of detection, identification and extraction on the KVR-1000 orthoimage were the main concern for this paper. RUSSIAN SPACE MAPPING SYSTEM KOMETA Fig. 1 shows the Russian Space Mapping System KOMETA equipped with TK-350 and KVR photographic cameras and the on-board devices including two star-trackers, laser altimeter and the positioning system for the determination of external parameters. Comet Class Spacecraft is launched periodically from the Baikonur Cosmodrome in Kazakhstan and orbit the Earth for about 45 days collecting up to 10 million square km of images. The entire satellite returns to Earth for film processing. Fig. 1: Russian Space System KOMETA equipped with KVR-1000 and TK-350 film-based cameras

3 EARSEL Workshop on Remote Sensing for Developing Countries, Cairo, In the case of high resolution KVR-1000 camera, exposure of the film is made by a moving slit passing in front of the photographic film. This gives rise to a cylindrical imaging surface (Fig. 2) on which each line is exposed from a different position in space with the possibility of varying tilt values occurring between them. As a result of the panoramic geometry, large differences in scale are experienced over the whole of the highly rectangular image format 18x72cm (40x160km on the ground). These large-scale changes have resulted in only the central 18x18cm part of the KVR-1000 photographic image being supplied to users in most cases in digital form after undergoing a prior scanning operation. Focal length of KVR-1000 camera is 1m and supplies the B/W panoramic imagery with the forward overlap of 6%-12% in the panchromatic band of of m. Ground resolution of such imagery is 1.56m with an image scale of 1:220,000. Cylindirical image Ground Flight direction Plane view Fig. 2: Panoramic distortion and its plane view. Other imaging system in KOMETA is TK-350 camera allows the acquisition of stereo-images during the same satellite pass with short time intervals. It has large film format of 300x450mm (200x300km on the ground) with a 350mm focal length. Main application of photographs from TK- 350 system in SIS is to produce 1: and smaller scale topographic maps in addition to the DEM generation for creating orthophotos from high resolution KVR-1000 images. Based on the publications of SIS s engineers, when no external ground control is used, the planimetric accuracy of these maps is typically 20-25m, and the vertical accuracy is 10m. If GPS derived control points are available, the accuracy of the maps increases to 15-20m horizontally and 5-7m vertically (Chekaline and Fomtchenko, 2000; Lavrov, 1996 and 2000). However, Buyuksalih et al. (2004) shows that the accuracy of about 10m in planimetry and 17m in height using the GCPs digitized from the 1:25000 scale topographic maps is indeed possible. The results draw a contradiction with the accuracy values found by SIS s engineers. EXPERIMENTAL AREA The test site is Zonguldak and its close vicinity, located in Western Black Sea region of Turkey. It is famous with being one of the main coal mining area in the world. Although losing economical interest, there are several coal mines still active in Zonguldak. Area has a rolling topography, in some parts, with steep and rugged terrain. While partly built city area is located alongside the sea coast, there are some agricultural lands and forests inner regions. The elevation ranges roughly up to 1800m. Fig. 3 shows the KVR-1000 orthoimage and the GCPs used in the accuracy evaluation phase. Given orthoimage is in the size of 8885x8752 pixels as a 8-bit greyscale GeoTIFF file referenced to the WGS84 ellipsoid and the UTM-zone 36 projection. Control points includes GPS surveyed points originally acquired for restituting high resolution Ikonos Geo-product imagery of the test area, the points from the aerial triangulation of 1:35000 scale aerial photographs of the test area and the points from the 1:1000 and 1:5000 scale maps. Such combination of different group

4 EARSEL Workshop on Remote Sensing for Developing Countries, Cairo, of points provide the better control point distribution. Digital image coordinates for these GCPs were measured manually using GCPWorks module of PCI system with sub-pixel point determination. Zonguldak City TURKEY Fig 3: Study area. In this test, image part from full panoramic KVR-1000 frame with frame number of 2252 and the viewing date of October 17 th, 2000 was implemented. As can be seen Fig. 4, Zonguldak testfield is lying close to the left edge of this KVR-1000 datatake in which panoramic distortion can also be expected. The first phase in the production of KVR-1000 orthoimages in SIS is the scanning of hardcopy KVR-1000 photographs. This task was realised by the Zeiss SCAI scanner using 7m pixel size. Fig. 4: Diagram of KVR-1000 frame location For rectification of KVR-1000 images, the PC-based digital photogrammetric system called Ortho/Z-Space developed by the cooperation of SIS and Russian Institute GosNIAS was used. In this process, generally DEM from stereo TK-350 images or by the available mapping materials can be used. In the given case, for ortho-image generation, DEM digitized from the topographic maps

5 EARSEL Workshop on Remote Sensing for Developing Countries, Cairo, of 1: scale (with the height accuracy of 20m) was used (information from the correspondence with Elena Molchinskina of SIS). GEOMETRIC ACCURACY TESTS Geopositional accuracy of KVR-1000 orthoimage was first analysed using the program Tran3D of Hannover University. In this program, user is able to run several transformations in sequence. First option in the program is the shift. For this, input to the program is the image and ground coordinates of GCPs and the header file of orthoimage. Using the WGS84 coordinates of upperleft corner of orthoimage and the pixel size supplied in the header file, image coordinates of GCPs were transformed to the given coordinate system. Then, program can compare the computed and given coordinates of GCPs and calculate the amount of shift to the specified coordinate system. This shift was m in X and m in Y. After correction of shift in each direction, rmse values were found to be ±19.60m in X and ±20.88m in Y-direction. This situation was improved by the similarity transformation which also respects the scale and rotation. Then, rmse-x and Y values were obtained as ±16.67m and ±15.36m respectively. However, application of affine transformation brought the further improvement in the accuracy values with ±13.46 in X and ±12.86m in Y. These are the points with higher elevation, indicating that the image orientation realised by SIS during the orthorectification of KVR image was not optimal. Actually, GCPs digitized from 1: scale maps were used in the orientation of Zonguldak KVR-1000 image (information from the correspondence with Elena Molchinskina of SIS). This can be corrected by the program CORIKON. The resulted rmse values for X and Y were equal to the ±10.20m and ±12.41m in Y respectively. As a result of the adjustment of the view direction, the error vectors of the elevated points are now in the usual range after adjustment with CORIKON. The data set was also adjusted with the satellite orbital modelling approach of PCI Geomatica system and produced nearly to the same result like achieved with CORIKON. Accuracy values of ±10.77m in X and ±12.61m in Y were acquired by this software. EVALUATION OF THE INFORMATION CONTENTS OF KVR-1000 IMAGES In order to extract information from the remotely-sensed image which is in raster form and has unique radiometric characteristics, in the first phase, image should be converted to the vector form. Although many methods are available for doing this step, most effective ways could be manual onscreen digitizing and object-oriented approaches. In this study, these two methods are used sequentially. Object-Oriented Analysis Object-oriented segmentation and following classification of KVR-1000 orthoimage was carried out using ecognition V4.0 software package (Definiens, 2004). The result is shown in Fig. 5. Then, different analyses have been realized by superimposing the segmented image onto the 1:1000 scale map in GIS environment (Fig. 6). During the experimental procedure, roads and buildings were tried to be extracted separately, but this was not the case because of the almost same grey values of both object groups. Best identification shown by the water structures, buildings, forest and green object classes. Based on the analysis made by MAPINFO 7.0 software using objectoriented approach, number of buildings extracted from the image remains as 752, it is equal to the 1059 in the case of 1:1000 scale maps.

6 EARSEL Workshop on Remote Sensing for Developing Countries, Cairo, Fig.5: Clasification of the test area from object-oriented segmentation. Fig.6: Building segmented image superimposed onto the 1:1000 scale map in GIS environment (black vectors from 1:1000 scale map, cells with red colour are result of object-oriented analysis) On-screen Digitizing On-screen digitizing of KVR-1000 ortho-image was realized by AutoCAD software package. Obtained vectoral information was compared with the 1:1000, 1:5000 and 1:25000 scale maps. For different information group, visual equality of the acquired on-screen digitizing results was compared to those derived from maps and given in Fig.8. In addition to this, the level of detectability and recognizability of the objects on KVR image were summarized in Table 1. As can be seen from this table, the main interpretation parameters are detectability and recognizability. According to STANAG 3769 (1998) the definition of detection is: In imagery interpretation the discovering of the existence of an object, but without recognition of the object, the definition of recognition is: The ability to fix the identity of a feature or object in imagery within a group type. Also in this table objects are classified based on the their level of detection and recognition. In this case, if object can be detected and recognized easily and sharply, then it is assigned as Perfect,

7 EARSEL Workshop on Remote Sensing for Developing Countries, Cairo, if the level was lower than, it is called as Good, if object only detected or recognized, then it is put into the class of Medium, if the level was very low, then it is Poor and the object is not available in the image, it is categorized as Not available (Samadzadegan, 2004). Apart from this, manually extracted buildings were superimposed onto the 1:1000 scale maps in GIS environment with different queries. While total number of buildings could be determined as 1059, only 78% of such buildings was extracted manually. Operator who undertakes the digitizing task has been considering the radiometric differences occur in the orthoimage while he/she was digitizing the objects manually. In this respect, contrast range and effective pixel size were ancillary parameters during the implementation. Gray value profil analysis has been used in the determination of effective pixel size. Two different constructions at some selected locations can be seen in Figure 7. At the constructions the edges depicted in green are selected. The aim of the following analysis is to determine contrast of edges, especially between white and dark areas. The dark areas can be the grass, water or ground, and the white areas can be road or roof of the buildings etc. The edge analysis consists of taking profiles along the edge itself. The profiles are chosen perpendicular to the edge. For all profiles, mean values of the corresponding profile points are determined. Thus a graphic showing the trend of average values is obtained. The differences of grey values between adjacent points are built and can be seen visually. Both graphics can be seen underneath the cropped samples in Figure 7. Based on the analysis made, effective pixel size for KVR-1000 image has been acquired as 2.7m (Şahin et all. 2004, Topan at all 2004). This shows that the operator was working with the 2 pixel not with 1 pixel which should be the case in the reality. Average * * * * * * * * * * * * * * * Differences 1-1 * 2-2 * 3-5 * 4-3 * 5 7 * 6 22 * 7 36 * 8 25 * 9 7 * 10 0 * 11 2 * 12 6 * 13 4 * 14 2 * Average * * * * * * * * * * * * * * * Differences 1-4 * 2-2 * 3 0 * 4 2 * 5-1 * 6 3 * 7 14 * 8 29 * 9 36 * * 11 5 * 12-1 * 13 0 * 14 1 * Fig.7: Grey value profiles of different edges. CONCLUSION In this study, manual and automatic object-oriented segmentation and classification performance of KVR-1000 orthoimage was tested. Therefore, results are collected under two different headings; For object-oriented automatic approach: Some objects cannot be extracted separately, for instance, roads-buildings and green areas with forest class. However, water structures can be identified successfully. Only 71 of the buildings can be extracted using an approach based automatic object-oriented classification. This is promising result for such imagery. In order to extract water structures, buildings and forest areas individually using ecognition software, scale factor should be assigned as 250. Edge of buildings cannot be detected sharply using object-oriented approach. On the other hand, buildings cannot be extracted with high success rate because they give same reflection profile with roads. Expected success rate could not be reached on the KVR-1000 ortho-image using ecognition 3.0 object-oriented image analysis software not enough contrast, monochromatic image, and negative influence of DEM on orthoimage generation fails the segmentation phase, then the following classification produce did not work as efficient as possible.

8 EARSEL Workshop on Remote Sensing for Developing Countries, Cairo, Table1: Summary of KVR image information content evaluation in different scale maps Detectability Manual Recognizability No Class Sub-Class Perfect Good Medium Poor Not Availble Excellent Good Medium Poor Not Availble 1 Road & Rail 2 Bulding 3 Area 4 Public Services 5 Water 6 Structure 7 Vegetation 8 Mining Structure Road Street Square Free Way High Way Bulvar Rail Building Block Single Building Mosque Parking Sport Building Medical & Educational Military Area Sports Area Camping Area Cemetry Park Industrial Area Waste & Coalwaste Area Power line Reservoir Pipe Line Tel Line Point Objects Line Objects Area objects Fence Tower Bridge Helipad Wall Dam Orchard Cultivate Forest Costline & Beach Lawn & Grass Single Tree Coal preparation plant Band Conveyor Sahft Tower Coal Washing pool ventilation shaft

9 EARSEL Workshop on Remote Sensing for Developing Countries, Cairo, For manual extraction: Experimental procedure was realized by the operator who has knowledge of the test field and capability of handling such imagery. When on-screen digitizing was carried out on KVR-1000, radiometric changes should also be taken into account. Analysis shows that effective pixel size of KVR-1000 image is about 2.7m. When compared to the automatic approach, manual extraction showed better performance in digitizing the highway and railways. Water structures can be detected and recognized sharply. Classification of vegetation can be realized only as grass and trees and deeper analysis is not possible with the available radiometric information. Different types of buildings can be detected and recognized in the medium level. From the GIS-based analysis, 78% of the buildings can be detected and recognized in the image. Theoretically, based on the ground pixel size of KVR image, it can stated that the mapping is possible with the scale of 1: However, practical works show that this image has more information contents than the one available in 1:25000 scale maps, in most cases, it can used as a base material in updating of 1:5000 scale maps. REFERANCES i. Buyuksalih G, M G Kocak, M Oruc, H Akcin, K Jacobsen, Accuracy Analysis, DEM Generation and Validation Using Russian TK-350 Stereo-images, Photogrammetric Record, 19(107). ii. Chekaline V F and M M Fomtchenko, Russian concept of the space images digital processing. International Archives of Photogrammetry and Remote Sensing, 33(4/1): iii. Definiens, ecognation user manual. iv. Doyle, FJ, Surveying and Mapping with Space Data A(4): , (ITC publication Series). v. Konency G, W Schrur, J Wu, Investigations of Interpretability of Images by Different Sensors and Platforms for Small Scale Mapping, Proceedings ISPRS Commission IV Symposium, Crystal City, vi. Li Z, High-resolution satellite images: past, present and future. Journal of Geospatial Engineering, 2(2): vii. Petrie G, Characteristics and applications of high-resolution space imagery. Mapping Awareness, 13(10): viii. Samadzadegan F, M Hahn, H Bagherzadeh & M Haeri, On the geometric accuracy and information content of ikonos high resolution imagery for map revision. In: XX. Congress of ISPRS, (İstanbul, Türkey). ix. Sovinformsputnik, Information from official web-site, (moscow, Russia). (accessed 18 Feb. 2004). x. Sahin H, H Topan, S Karakış & A M Marangoz, Comparison of object oriented image analysis and manual digitizing for feature extraction, XX. Congress of ISPRS, (İstanbul, Turkey). xi. Topan H, G Büyüksalih, K Jacobsen, In: Comparison of İnformation contents of high resolution space images, Int. Arch. of Photogrammetry and Remote Sensing, 34 (B?), Istanbul, Turkey, ACKNOWLEDGEMENTS This study is supported by TUBITAK, Turkey and the Julich Research Centre, Germany. The authors wish to thank their German partner Dr. Karsten Jacobsen for his suggestions.

10 EARSEL Workshop on Remote Sensing for Developing Countries, Cairo, KVR & 1:1000 scale map 1:5000 scale map 1:25000 scale map Building blocks ( Green lines from KVR-1000 and red line from 1:1000 scale map) Single buildings (white lines from KVR-1000 and red line from 1:1000 scale map) hospital Fig.8: On-screen digitizing of KVR-1000 ortho-image with compared to those derived from different maps

11 EARSEL Workshop on Remote Sensing for Developing Countries, Cairo, University Buildings Sport complex Bridges Fig.8 Continued.

12 EARSEL Workshop on Remote Sensing for Developing Countries, Cairo, Port Industrial area (coal preparation plant) Coal preparation plants Fig.8 Continued.

13 EARSEL Workshop on Remote Sensing for Developing Countries, Cairo, Mosque (with dome) River lying in the canal Dam Roads (with their central lines) Fig.8 Continued.

14 EARSEL Workshop on Remote Sensing for Developing Countries, Cairo, Roundabout Railway type 1 Railway type 2 Fig.8 Continued.

15 EARSEL Workshop on Remote Sensing for Developing Countries, Cairo, Separation of rocky, grassy and forest parts Coastal line Helipad Fig.8 Continued.