LAND COVER / LAND USE MAP OF GERMANY BASED ON MERIS FULL RESOLUTION DATA U. Geßner (1), K. P. Günther (2), S. W. Maier (3) (1) German Aerospace Center (DLR) German Remote Sensing Data Center (DFD) Münchnerstraße 20, D 82230 Weßling ursula.gessner@dlr.de (2) German Aerospace Center (DLR) German Remote Sensing Data Center (DFD) Münchnerstraße 20, D 82230 Weßling kurt.guenther@dlr.de (3) Department of Land Information Satellite Remote Sensing Services 65 Brockway Road, Floreat, WA 6015, Australia stefan.maier@dli.wa.gov.au ABSTRACT This paper describes first results of a land cover / land use (LCLU) of Germany, developed at the German Remote Sensing Data Center (DLR-DFD). The automated, yearly updated product is based on MERIS level 2 full resolution data. The procedure consists of two main steps, namely a multispectral and a multitemporal analysis. The legend of the LCLU map is defined according to the system LCCS developed by FAO and UNEP. The new product is dedicated indicating land cover / land use changes at medium spatial resolution which are of importance in research areas such as climate change. At a later date, the spatial extension of the map to the entire area of Europe will be performed. 1. INTRODUCTION The land cover / land use situation of Europe is changing continuously due to both human activities and environmental, mainly climatic, conditions. Detection, monitoring and mapping of LCLU are important issues for environmental sciences and play a significant role in fields such as climate modelling, carbon cycle studies, agricultural purposes and resource management. Many approaches to mapping LCLU are based on remote sensing with earth observation satellites. These approaches make use of the advantage of continuous monitoring in space and time. Some studies use fine resolution data of pixel sizes smaller than 100m x 100m. These approaches usually result in precise maps of limited areas. An example of a fine resolution land cover that is performed over more expanded areas is the CORINE land cover project [2]. This project is based on timeconsuming visual interpretation of Landsat-TM data for the whole of Europe, which makes an update feasible only in time steps of several years. Other land cover mapping approaches are based upon coarse resolution data. In this context, data from sensors like NOAA- AVHRR, with a spatial resolution of 1km, allow land cover mapping at a global scale [5/7]. Medium resolution sensors like MERIS or MODIS play an intermediate role concerning the spatial resolution and the dimension of the mapping area. Land cover mapping approaches using MERIS data are mostly performed at regional or national scales [3/10]. Furthermore, LCLU mapping approaches can be differentiated by a temporal aspect. Monotemporal studies use satellite data of only one single point in time [6], whereas many other approaches analyse several images - usually a limited number of selected scenes. These multitemporal methods take advantage of the phenology of vegetation [9] or they are in particular interested in change detection [10]. Within the ESA-AO project GEMEL3 (GEneration of MERIS Level 3 products for European Multidisciplinary Regional Applications, AO-ID 1413), an automated LCLU of Germany based on MERIS full resolution (300m x 300m) data is developed. In a multispectral and a multitemporal step, this analyses the complete MERIS data record set of one year. 2. DATA Within the scope of the project GEMEL3, the German Aerospace Center (DLR) in Neustrelitz has been receiving MERIS full resolution data nearly continuously since July 2003. The data is processed using the ESA standard processor (version 4.0) and archived for further use at the DLR. The is based on MERIS level 2 data in full resolution with 300m x 300m pixel size. The sensor is Proc. of the 2004 Envisat & ERS Symposium, Salzburg, Austria 6-10 September 2004 (ESA SP-572, April 2005)
composed of five CCD-arrays and has a swath width of 1150km. It delivers images of Central Europe of 1 to 2 orbits per day. MERIS level 2 data consists of geolocated, calibrated surface reflectances ( top of aerosol ) in 13 spectral channels from visible to near infrared wavelengths (412.5 to 890nm). For the calculation of the annual multitemporal LCLU, the entire MERIS data record set of one year is used. 3. CLASSIFICATION PROCEDURE The automated procedure is based on two main steps. First, a supervised multispectral is performed using level 2 data in sensor projection. After map projection the multispectrally classified data is analysed multitemporally. The structure of the LCLU processor is shown in fig. 1. The supervised calculates only four classes, one for dark pixels such as land and water, one for thin clouds over land, one for thin clouds over water and a last one for thick clouds. First results of this cloud identification are shown in fig. 2a/b. The method has rather a tendency to reject a not-clouded pixel from than to admit a thin-clouded pixel. Though some valuable pixels might be neglegted, most importantly, erroneous admission of clouded pixels to is minimised. In view of the large amount of data collected yearly and the negative influence of cloud pixels on results, this tendancy is considered an acceptable trade-off. Segmented spectral data base statistics Maximum-likelihood viewing/illumination geometry ESA-Level 2 top-of-aerosol reflectances ESA-Level 1, 2 flags (sensor projection) Level 2 colour class (sensor projection) Fig. 2a. Original MERIS full resolution image Map projection Multitemporal data base scheme multitemporal geographical coordinates Level 3a colour class (map projection M1 Grid) Level 3c land cover class (map projection M1 Grid) Fig. 1. Structure of the MERIS LCLU processor 3.1 Preparation of data Pixels representing clouds, water or invalid values are removed from further calculations. Until now, flags of MERIS L1 and L2 products were used to identify these pixels. Internal investigations have shown that the level 2 cloud flag frequently does not assign all thin clouded pixels, as this flag indicates the availability of cloud products. As thin clouds worsen the result, it is tested to identify clouded pixels by means of a different method, namely a maximum likelihood. Fig. 2b. Result of cloud identification by means of a supervised. Thick clouds are marked in yellow, thin clouds in blue. 3.2 Multispectral step In the main procedure, as a first step, a supervised maximum likelihood is performed for each available MERIS level 2 image of one year. MERIS level 2 data is not aerosol-corrected over land and channels 1 to 3 (412.5 to 490nm) show a strong influence of atmospheric conditions. For this reason, the multispectral step omits these
wavelengths. Potential further reduction of channels is being investigated. To enable an automated process, the supervised is performed by means of a data base which can be reused every year. This data base contains all statistical parameters that are needed for calculation, like the mean spectra, the inverse covariance matrices and the logarithm of the determinant for every class. Due to the sensor s exceptionally large swath width of 1150km (68.5 ), the bidirectional reflectance distribution influences the characteristics of spectral signatures measured by MERIS. Reflectances change according to the particular combination of sun zenith, view zenith and azimuth difference angle between sun and sensor. Fig. 3a and 3b show examples of these BRDF effects. The influence of backscatter and forward scatter conditions on the spectral signature of a grassland area is shown in fig. 3a. The spectra are extracted from two scenes in October with comparable sun zenith and view zenith angles. In both cases, the view zenith angles are greater than 30. In fig. 3b., the effect of changing view zenith angle is demonstrated for an evergreen needle forest in June. The three spectral signatures were measured under comparable azimuth difference and sun zenith angles, but varying view zenith angle conditions. To consider the differences in spectral signature caused by BRDF effects, the data base is segmented according to illumination and observation angles. For every combination of three sun zenith angle ranges, three view zenith angle ranges and two azimuth difference angle ranges, the corresponding statistical parameters for every class are stored separately in the data base. Due to this segmentation, there is no need for a seperate BRDFcorrection in the preprocessing stage. The segmented spectral data base is stored as an hdf5-file and filled by extracting typical spectral signatures from representative and homogeneous training areas located in Central Europe. The multispectral delivers three results. The first result corresponds to the first most likely class as demonstrated by the shortest malahanobis distance (called class 1) and the second result corresponds to the second most likely class (called class 2). The third result describes the ratio of the malahanobis distances of class 1 and class 2. This ratio can be interpreted as a separability index or a quality assessment. In case the separability index is close to zero, the membership of the regarded pixel to the first class has highest probability. The assignment to class 1 or to class 2 is ambiguous if the separability index goes to one. Later, these three results of the multispectral step are used in the multitemporal analysis, resulting in a fuzzy algorithm. Fig. 3a. Spectral signatures of grassland in October, observed under backscatter (56.02 ) and forward scatter (120.5 ) conditions. The view zenith angles are >30. dark green green brown blue unclassified Fig. 3b. Spectral signatures of evergreen needle forest in June, observed under backscatter conditions and three different view zenith angles (5.76 /14.97 /36.16 ). light green brown-green Fig. 4. Multispectral result for parts of southern Germany in August.
An important aspect of the monotemporal, multispectral step is the fact that the resulting classes are not yet LCLU classes but are instead intermediate so-called colour classes (see legend fig. 4). This approach assumes that different LCLU classes can have similar spectral properties at any given time. As a result of this first step, pixels belonging to different LCLU classes can be assigned to the same colour class, when their stages of phenology at a given time are similar. Our multispectral results showed differentiation between the five CCD-sensors of MERIS. Visually homogeneous land cover areas were sometimes identified as two different classes when their coverage spanned two adjacent CCDs. 3.3 Multitemporal step Most LCLU classes can not be distinguished easily when they are examined only at a single point in time, as is done in the first, multispectral step. As soon as the phenology of vegetation throughout the year is taken into account, the assignment to LCLU classes becomes much more evident. For this reason, in a second step, the multispectral results are analysed multitemporally. The multitemporal examination is based on the comparison of so-called reference vectors with pixel vectors. A reference vector describes the ideal temporal sequence of colour classes for a certain LCLU class. At least one reference vector must be built for every LCLU class. Reference vectors are defined by analysing the change of colour classes of the training areas during the year and with the help of phenological observations of the German Meteorological Service (DWD). With respect to variations in phenology, regional reference vectors have to be built, especially for a later extension of the LCLU map to the entire area of Europe. A pixel vector contains the actual temporal sequence of colour classes (class 1 and 2) at one location. This sequence is derived by examination of all multispectrally classified MERIS scenes of one year. For every pixel, the pixel vector is compared with all reference vectors, considering the three results (class 1, class 2 and separability factor) of the monotemporal step. In case class 1 or class 2 of one element of the pixel vector fits to the corresponding element of a reference vector, the separability index expresses the degree of agreement. The pixel is assigned to the LCLU class with the highest number of matches. In case a certain minimum of matches is not reached or in case two classes show the same number of matches, the pixel is not classified. Only after this second, multitemporal step, pixels are assigned to the final LCLU classes. Fig. 5. illustrates the procedure of multitemporal analysis. Phenological data (DWD) Analyse of training areas Reference vectors Cropland Grassland Pixel vector Deciduous forest Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Fig. 5. Schematic illustration of the multitemporal step A pixel vector is compared with reference vectors for cropland, grassland and deciduous forest and then assigned to the class with the most matches (in this case cropland). All pixels which were classified in the multispectral step can be integrated in this multitemporal analysis. The data has to be mapped to a common projection, but there is no need to mosaic in connection with this method. In order to identify pixels showing exactly the same spot on the surface in different scenes, a precise geocoding of data is necessary. Unfortunately, the geocoding of MERIS L2 does not fulfil this requirement for every image at the moment. In several scenes, deviations of more than 4 pixels (1.2km) occur; in extreme cases even deviations of 30 pixels (9km) were detected. Geometric inaccuracies of this dimension make an automated multitemporal analysis impossible. As every available image of the mapping area will be analysed, it is not feasible to perform an additional geometric correction with the aid of ground control points. A potential procedure to correct the geometric accuracy of MERIS full resolution data might be realised with the help of GEOCOL, a tool of the software package XDIBIAS, developed at DLR [8]. GEOCOL identifies homologous points with the result that two scenes can be compared. This possibility still has to be checked. 4. OUTLOOK At present, the realisation of the LCLU is in progress. First results show good assignment of the pixel vectors to colour classes by means of the supervised maximum likelihood. About 400 MERIS full resolution scenes are in the monotemporal processing loop. Problems appeared mainly concerning the multitemporal analysis because of the quality of geolocation of several MERIS scenes. Additionally, in respect of discrepancies of the five CCD-sensors, ----- - Number of Matches 8 3 6
difficulties in the multispectral step appear. Further work will concentrate on the optimisation of the multispectral procedure concerning for example the amelioration of threshold values and the reduction of the number of regarded channels. That followed, the multitemporal step can be realised, when the geolocation precision has improved. Once the final has been completed and optimised, it is planned to estimate accuracy using GLC2000. Before this can happen, the GLC2000 data must be aggregated to the legend and the spatial resolution of the MERIS LCLU map. Following the production of a LCLU map of Germany, the procedure is planned to be extended to the entire area of Europe. For this purpose, slight modifications, especially concerning the multitemporal analysis, will be necessary. Primarily designed for application in ocean analysis, MERIS has proven to be a suitable instrument for the investigation of land surfaces as well. 5. REFERENCES 7. Loveland T. R., et al., Development of a global land cover characteristics database and IGBP DISCover from 1-km AVHRR data, International Journal of Remote Sensing, Vol. 21, 1303-1330, 2000. 8. Müller R., Reinartz P., Kritikos G. and Schroeder M., Image Processing in a Network Environment, Proceedings of ISPRS 92, Washington, Internat. Archives of Photogrammetry and Remote Sensing, Vol. 29, 289-293, 1992. 9. Oetter D. R., Cohen W. B., Berterretche M., Maiersperger T. K., Kennedy R. E., Land cover mapping in an agricultural setting using multiseasonal Thematic Mapper data, Remote Sensing of the Environment, Vol. 76, 139-155, 2000. 10. Seiler R. and Csaplovics E., Monitoring landcover changes of the Niger inland delta (Mali) by means of Envisat-MERIS data, http://envisat.esa.int/workshops/ meris03/participants/112/paper_meris-workshop_ 2003_finalversion.pdf, 2003. 1. Arndt M., Landbedeckungskartierung von Deutschland auf der Grundlage einer optischen Satellitendaten Zeitreihe: In Vorbereitung des MERIS- LCC-Prozessors, Diplomarbeit, Universität Köln [unpublished], 2001. 2. Bossard M., Feranec J. and Otahel J., CORINE Land Cover Technical Guide Addendum 2000, European Environmental Agency, Technical Report No. 40, 2000. 3. Clevers J. G., Bartholomeus H. M., Mücher C. A. and de Wit A. J., Use of MERIS data for land cover mapping in the Netherlands, http://envisat.esa.int/ workshops/meris03/participants/22/paper_cleverspaper-meris-workshop2003.pdf, 2003. 4. Günther K., Neumann A., Gege P., Doerffer R., Fischer J. and Brockmann C., MERIS Value added products for land-, water- and atmospheric applications, in: Dech S., et al. (Ed.), Tagungsband 19. DFD- Nutzerseminar, 15. 16. Oktober 2002, 37-51, Oberpfaffenhofen, 2002. 5. Hansen M. C., et al., Global land cover at 1km spatial resolution using a tree approach, International Journal of Remote Sensing, Vol. 21, 1331-1364, 2000. 6. Keuchel J., Naumann S., Heiler M. and Siegmund A., Automatic land cover analysis for Tenerife by supervised using remotely sensed data, Remote Sensing of the Environment, Vol. 86, 530-541, 2003.