REVISION OF TOPOGRAPHIC DATABASES BY SATELLITE IMAGES Bettina Petzold Landesvermessungsamt Nordrhein-Westfalen Muffendorfer Str. 19-21, 53177 Bonn Tel.: 0228 / 846 4220, FAX: 846-4002 e-mail: petzold@lverma.nrw.de Volker Walter Institut für Photogrammetrie, Universität Stuttgart Geschwister-Scholl-Str. 24, 70174 Stuttgart Tel.: 0711 / 121 4091, FAX: 121-3297 e-mail: volker.walter@ifp.uni-stuttgart.de KEYWORDS: Updating GIS, High Resolution Imaging Satellites, Laser Data ABSTRACT: This paper examines data from different sensors regarding their potential for an automatic revision of topographic databases. The data which have to be updated are from the German national topographic cartographic database (ATKIS) and were captured in the scale of 1 : 25,000. After a brief introduction into ATKIS the used approach is discussed. Results are shown on examples of data from several sensors: scanned analogue aerial photos, an airborne digital line scanner (DPA camera system), the Indian satellite IRS-1C, the MOMS-2P camera and from a laser scanning system as an additional information source. 1.1 The Basics 1. ATKIS DATA ATKIS, the Authoritative Topographic-Cartographic Information System (Amtliches Topographisch-Kartographisches Informationssystem) is one of the common projects of the Federal Republic of Germany State Survey Working Committee (Arbeitsgemeinschaft der Vermessungsverwaltungen der Länder der Bundesrepublik Deutschland (AdV)). This committee was established to achieve a harmonized development within surveying and mapping which are in Germany within the responsibility of the Federal States. The ATKIS data model is object structured, uses attributes on several levels, allows hierarchical relations between objects and enables the aggregation of objects to complex objects. The data are stored in vector format. 1.2 The Object Class Catalogue and The Digital Landscape Model In the Object Class Catalogue (Objektartenkatalog (OK)) all landscape objects are specified by their definition, their type (point, line or area) and the definition of the accompanying attributes and their possible values. ATKIS objects can be visible real objects of the landscape but also administrative objects as districts and boundaries. The objects are grouped in object classes, subsequently in object groups and finally in object domains. The digital landscape model (Digitales Landschaftsmodell (DLM)) DLM 25 covers the scale domain of about 1:25.000 and comes up to the most urgent requirements of most part of the users. That is why it has the highest priority, it is called the basis DLM and is the first to be built up by the Surveying and Mapping Agencies (Landesvermessungsämter (LVermA)) of the Federal States of Germany. A first version of the DLM 25, the so-called DLM 25/1, consists of the most important object classes such as the networks of roads, railways and waters and all covered areas between these network lines. 1.3 Data Capture In Northrhine-Westphalia (NRW) the German Base Map 1:5000 (Deutsche Grundkarte 1:5000, DGK 5)) and the Orthophoto Map 1:5000 (Luftbildkarte, DGK 5L) are used as the main source of information. They guarantee the required precision of ± 3 m and contain the information for capturing of most of the defined attributes. In addition some special small scale maps are necessary, containing attribute information like the type of main roads, of railroads and the river network or some information of power providing companies. A complete coverage of the landscape is guaranteed by a defined group of object classes.
2.1 Definitions 2. UPDATING Updating is defined as a systematic revision of the ATKIS database. The updating can be subdivided into the tasks of detecting of field changes, procuring of necessary information and incorporating of new information into the existing database. The procuring of the necessary information is defined as the obtaining of information extensive enough for updating. That means the geometry has to have the required precision, and the semantic information has to be complete. The distinction between detecting and procuring is necessary because only some of the possible methods of data acquisition are supporting both parts. Regarding satellite data it directly depends on the ground resolution whether those data can be used to detect changes only or in addition to procure the information necessary for updating. Finally the incorporation is the introduction of the acquired information into the ATKIS-database. 2.2 Information Sources old and new The field check and the topographic information service have the oldest tradition as they have already been used for the updating of analogue topographic maps. But the working methods for the updating of ATKIS have to be changed because there are much more information in ATKIS than in analogue maps and the data have to be updated continuously and not only periodically. Low or medium spatial resolution satellite images are already used for the updating of geographic information systems in scales of 1:50.000 or smaller, for example SPOT or LANDSAT data [Konecny 96]. The ground pixel size of 10 m or more prevents those data from fulfilling the requirements for the updating of ATKIS. More interesting could be the data of the Modular Optoelectronic Multispectral Stereo Scanner (MOMS) which is installed on the PRIRODA module of the Russian MIR station. MOMS has a high resolution panchromatic sensor with a ground pixel size of about 7 m. First MOMS data have been scheduled for the beginning of 1996 and were the first possibility to test a remote sensing product with a good spatial resolution, even though not sufficient for the updating of ATKIS. New high resolution satellite systems are presently under development and will be operational soon - thus high resolution remote sensing data will also be available soon [Jacobson 98], [Kilston 98] which will have a resolution which should be sufficient for the update of ATKIS. 3.1 The General Idea 3. PROJECT In spring 1994 the German Space Agency (DARA), today the German Aerospace Centre (DLR), called for proposals for special exemplary projects using MOMS data for administrational tasks. A financial support should be given for those projects which should help to contribute to the exploration of new fields of application of remote sensing. The LVermA NRW took the opportunity, and finally the project was supported by the DLR. The objectives of the project are formulated as follows: - an automatic approach for the detection of temporal changes in ATKIS based on satellite orthophotos has to be developed; - as a result of the project the information was expected which object classes can be updated by the developed approach and under which conditions. The objective is not only to update the geometry but, if possible, also the attributes. The strategy for the updating should fit especially into MOMS data, but also to images with a higher spatial resolution up to already existing scanned analogue orthophotos. That is done first because satellites with a high spatial resolution are already announced [Fritz 97]. And second it should be sure that the algorithms would even fit into orthophotos for the worst case that no up-to-date satellite images could be used, for example after a long bad weather period. The developed approach should help at least in the most tiring and hence the most error-prone part of updating, the task of detecting the field changes by manual revision of orthophotos. 3.2 The Approach The approach is fully automatic and can be subdivided in general into two steps (see figure 1). First the data are classified with a supervised Maximum Likelihood Classification, then the classification result has to be compared with existing ATKIS objects. In the following more general details of the process are given whereas some test results are discussed in chapter 4.
sensor data (multispectral + preprocessed channels) training areas derived from GIS database supervised maximum likelihood classification GIS data pixel oriented classification result matching of classified data and GIS database revised GIS data Figure 1: Overview of the automatic verification of ATKIS 3.2.1 The Training Areas and Classification The training areas are not digitized manually by an operator but they are derived automatically from the existing ATKIS data. This is possible if the number of field changes or possible restitution errors is considerably lower than all ATKIS objects. For more details of the generation of training areas see [Walter 98]. Not all ATKIS object classes can be distinguished only by spectral or textural specifications. Therefore the ATKIS object classes are grouped together into six land use classes. These are forest, agricultural areas, garden, settlement, water and street. In addition the class shadow can be defined. For the land use class settlement data with low resolution (e. g. > 15 m) can be used as they are. But data with higher resolution necessitate special preparations: the training areas should consist only of those pixels that represent buildings. To achieve this, a buffer around streets is computed and pixels which represent vegetation are deleted with the help of the vegetation index. The remaining pixels are most probably buildings. It only makes sense to define a land use class street if the ground pixel size is of 3 to 4 m or less. Otherwise the pixels are more or less mixed pixels leading to no result. For the training areas only those pixel laying on the middle axis of the object are used, and vegetation pixels are again deleted. In addition, all streets laying in forests are not used for the training areas. 3.2.2 Matching After the classification it must be decided which of the GIS objects do not match the classification result. This can be objects where a change in the landscape has occurred or objects, that were not collected correctly. All GIS objects are subdivided into three classes. The first class contains all objects which could be detected with a high certainty, the second class contains all objects which are detected only partly and the third class contains all objects which could not be detected at all. The decision to which class an object belongs is made by measuring the percentage of pixels which are classified to the same object class as the object in the GIS database belongs to. Besides the percentage of correctly classified pixels, also the homogeneity of the classification result and the form are used for verification. A detailed description of the
DPA & ATKIS classification matching a) greenland b) settlement streets settlement greenland forest full verified partly verified not found Figure 2: Examples for classification and matching of DPA data verification of the classification results can be found in [Walter 1988], [Walter, Fritsch 98]. Figure 2 shows some results of the verification of ATKIS objects in DPA images. The DPA images are represented in the original resolution in the left column. The ATKIS geometry is superimposed in black. The classification result in a resolution of 2 m is represented in the middle column and the result of the verification in the right column. Figure 2 a) shows two objects which were captured in ATKIS as greenland. In the DPA image it can be seen that meanwhile a settlement area was built up. The result of the verification is that these two objects cannot be found in the image because of the low number of pixels that were classified as greenland. Figure 2 b) shows an settlement objects which contains a big greenland area. The object is marked as partly found because at least the left part of the object was classified as settlement or streets. 4. TESTS WITH DIFFERENT DATA SETS When starting the project no MOMS scene of Northrhine-Westphalia was available. That is why other data were tested first. 4.1 IRS Data As more and more problems with the MIR station occurred, the DLR decided to provide all participants of MOMS projects with IRS images. The IRS-1C consists of a high resolution panchromatic sensor (PAN) with a spatial resolution of 5.8 m and a radiometric resolution of 6 bit, a sensor with four multispectral channels with a radiometric resolution
IRS-1c (cir) DPA classification ATKIS water forest settlement green land Figure 3: Classification of IRS-1C data of 7 bit and a spatial resolution of 23.5 m and finally a wide field sensor (WiFS) with a spatial resolution of 188 m which is without any interest for our purposes. More details about the Indian Remote Sensing Satellite (IRS) 1C can be found in [Srivastava 96]. The expectations concerning those data were not too high because of the low spatial resolution but the data enabled a test of the algorithms and a first confirmation that the basic idea is correct. Figure 3 shows a classification result of IRS-1C data. The image is shown as a CIR image because IRS-1C does not capture data in the blue range. In order to improve the interpretation of the IRS-1C image, a DPA image of the same area is represented. The results confirmed that the developed algorithms are working correctly. But the further processing of the data is very problematic. If an object is marked as not found it would be very difficult for an operator to decide if there has been a change or not if he has only IRS-1C data as a decision base. 4.2 MOMS Data The German MOMS sensor (Modular Optoelectronic Multispectral Stereo Scanner) is installed on the PRIRODA module which is docked to the Russian MIR station. The scanner consists of a high resolution panchromatic 3-line stereo module and a multispectral camera with a red, green, blue and near infrared channel. The ground resolution is 7 m for the panchromatic nadir channel, 15 m for the forward and backward looking channel and 18 m for the multispectral data. For more information on the MOMS camera see [Schiewe 98]. A pixel precise classification is partially possible here. But problems appear for example between the differentiation of the land use classes settlement and forest. The explanation for these problems was found quickly: the radiometry of that MOMS scene was very bad, in all channels less than 100 grey values are used. One detail of the classification result can be seen in Figure 4. 4.3 DPA Data DPA is a digital line array scanner system with four multispectral channels (red, green, blue, and near infrared) and a three-line panchromatic stereo module, very similar to the MOMS camera but developed for airborne applications. The radiometric resolution is 8 bit for all channels. The ground pixel size directly depends of the flight height above ground. A flight height of 1900 m results for example in a ground pixel size of 60 cm for the multispectral channels and 30 cm for the stereo images.
MOMS-2P (cir) classification orthophoto (rgb) ATKIS other water forest settlement green land Figure 4: Classification of MOMS-2P data The data of the test area have a ground pixel size of 0.75 m. This corresponds at an area of 2 km * 2 km to a pixel number of more than seven million pixels. Because the classification is a complex process, this leads to a high computing time as well as high memory requirements. However, experiences show that depending on the land use class high resolution must not necessarily improve the results of the classification. In order to find a compromise between quality of change detection and computer requirements the data were resampled to a pixel size of 2 m. Examples of the classification of DPA data are shown in Figure 2. 4.4 Usage of Laser Data Because the software was developed in such a way that there exist no restrictions regarding the input data for the classification it is also possible to use laser data as an additional information channel. First we tested the combination of standard rgborthophotos and laser data. But it had to be stated that even with the known height information derived from laser data the result was very week in some parts of the image. Especially in shadow regions the classification was bad. That can be stated especially for the land use class street. The spectral characteristics of street pixels laying in the sun are very different from those laying in the shadow. Another test was the classification of cir orthophotos with and without laser data. The laser data improve the classification result significantly because they have a complementary behaviour as multispectral data. With laser data the classes agricultural areas and street can be separated very good from the classes forest and settlement because of the different heights of the pixels above the ground whereas in multispectral data the classes farmland and forest can be separated very good from the classes streets and settlement because of the strongly different percentage of chlorophyll. One classification result is shown in Figure 5. Shadow areas can be classified considerably better in cir images than in rgb images. In areas with very strong shadow it comes also to wrong classifications in cir images, however considerably less than in rgb images. Our next strategy was to introduce an additional land use class shadow in the classification process. Because there is no information in the GIS database about shadow areas, they have to be derived from the laser data. For the automatic generation of training areas for shadow the local height which is provided by the laser data as well as the elevation and azimuth of the sun at the time of image acquisition is required. The elevation and azimuth of the sun can either be determined manually by an interactive measurement of the edge of a shadowed area in the image and the
orthophoto (cir) classification without laser data classification with laser data forest greenland settlement streets Figure 5: Classification without and with laser data corresponding object height in the normalized laser data or derived automatically from the geographical latitude and longitude of the captured area and the time of image acquisition. Figure 6 shows the automatic processed training areas for the additional land use class shadow. In order to avoid the shadow class in the final result the approach can be further refined by splitting each of the land use classes into one land use class for shadow areas and one land use class for non shadow areas. After the classification the shadow and non shadow pixels for each land use class are combined again to obtain one unique class for each type of land use. The final result of the classification algorithm for the whole test area is given in Figure 7. 5. FINAL RESULTS AND CONCLUSIONS In this paper the potential of data from different sensors for the verification of ATKIS is examined. Data from the sensors IRS-1C and MOMS-2P are only to a limited extent suitable for this approach because of their low resolution. Objects must have a size of 2 * 2 pixel so that they can be recognized certainly in the figure. A recognition of line objects, like streets, is not possible. A big problem is the verification of the classification results. Even if objects can be recognized pixel precisely, it is not possible for an operator to verify the results without further information sources. With data from the DPA camera system, good results can be achieved. Area objects can be recognized in a sufficient accuracy for ATKIS. Figure 6: Automatic generation of training areas for shadow
orthophoto classification water forest greenland settlement streets Figure 7: Classification using 10 classes (5 for shadow areas and 5 for non shadow areas) However, in inner city areas street pixels may be classified as houses and vice versa. This is a problem for the road detection. A reliable detection of roads with DPA data is possible only in sparsely populated areas. The DPA data were resampled to a resolution of 2 m. This falls into the range which will be covered by future high resolution satellite systems (see for example [Jacobsen 98] or [Kilston 98]). With the availability of such systems an automated verification of data in the scale 1:25,000 based on up-to-date data will be possible. The tests showed that it is not possible to achieve homogenous good classification results based on rgb data. Even the addition of laser data does not lead to sufficient results. A channel in the nir range is necessary to handle shadowed areas. The best results were achieved with the combination of cir images and laser data. With this combination it is also possible to verify objects in larger scales down to 1:2.500 (for example ground plans of buildings) if the images and the laser data are captured in a high resolution. This is also the topic which we want to focus our future work on. A problem is the definition of quality measures to compare the performance of the different sensors. The classification result can be described with statistical measures such as the Kappa value or other measures described in the literature [Congalton 91] [Rosenfeld and Fritzpatrick-Lins 86] [Stehmann 97] but it is difficult to define a quality measure for the object verification. Problems especially appear with objects that are captured according to ownership structures and not to detectable structures in the image. Objects of this kind often have inhomogeneous spectral and textural characteristics and even a human operator is not able to decide if there is a change in the landscape or not if he has no additional information source. Also different operator will come to different decisions. Another problem is that some object classes are defined ambiguously and therefore are not clearly delimitable from each other. In order to get a more practical oriented idea of the quality of the results we will install the developed software package at the Surveying Institute of the State of Northrhine- Westphalia. Extensive data sets will be processed there and the results will be evaluated by ATKIS professionals. 6. REFERENCES Congalton, R. 1991: A Review of Assessing the Accuracy of Classifications of Remotely Sensed Data, Remote Sensing Environment 37, 35 46 Fritz, L. W.: August 1997 Status of New Commercial Earth Observation Satellite Systems; Photogrammetrie, Fernerkundung, Geoinformation 6/1997, p. 369-382 Hahn, M., Stallmann, D., Staetter, C. 1996: The DPA-Sensor System for Topographic and Thematic
Mapping; in: International Archives of Photogrammetry and Remote Sensing, Vol XXXI, Part B2, p. 141-146 Jacobsen, K. 1998: Status and Tendency of Sensors for Mapping, Proceedings of the International Symposium on Earth Observation System for sustainable Development, Bangladore, India, Vol. XXXII, Part I of International Archives of Photogrammetry and Remote Sensing (ISPRS), 183 190 Remote Sensing (ISPRS), Vol XXXII, Part 3/1, p. 485-489 Walter, V. 1999; Comparison of the potential of different sensors for an automatic approach for change detection in GIS databases; in: International Workshop on Integrated Spatial Databases: Images and GIS, accepted for publishing. Kilston, S. 1998: Capabilities of new Remote Sensing Satellites to support sustainable Development, Proceedings of the International Symposium on Earth Observation System for sustainable Development, Bangladore, India, Vol. XXXII, Part I of International Archives of Photogrammetry and Remote Sensing (ISPRS), 124-131 Konecny, G. 1996: International Mapping from Space; International Archives of Photogrammetry and Remote Sensing, Vol. XXXI, Part B4, Vienna Petzold, B. 1998; High Resolution Satellite Images For The Updating Of Geotopographic Basis Information; in: Wissenschaftliche Arbeiten der Fachrichtung Vermessungswesen der Universität Hannover Nr. 227, Festschrift Univ.-Prof. Dr.-Ing. Dr.h.c. mult. Gottfried Konecny zur Emeritierung, p. 243-252; Schiewe, J.: MOMS-02: Gelungenes Experiment ohne Zukunft?; Photogrammetrie, Fernerkundung, Geoinformation 1/1998, p. 17-25 Srivastava, P. 1996; Cartographic Potential of IRS- 1C Data Products; in: International Archives of Photogrammetry and Remote Sensing, Vol. 31, Part B4, page 823-828 Stehmann, V. 1997: Selecting and Interpreting Measures of Thematic Classification Accuracy, Remote Sensing Environment 62, 77-89 Walter, V. 1998; Automatic classification of remote sensing data for GIS database revision; in: ISPRS Commission IV Symposium on GIS Between Visions and Applications, Stuttgart, Germany, p. 641-648 Walter, V., Fritsch, D. 1998; Automatic verification of GIS data using high resolution multispectral data; in; International Archives of Photogrammetry and