Hydrologic land cover classification mapping at local level with the combined use of ASTER multispectral imagery and GPS measurements

Hydrologic land cover classification mapping at local level with the combined use of ASTER multispectral imagery and GPS measurements Nektarios Chrysoulakis* 1, Iphigenia Keramitsoglou 2 and Constantinos Cartalis 2 1 Foundation for Research and Technology Hellas, Institute of Applied and Computational Mathematics, Regional Analysis Division P.O. Box 1527, GR-71110, Heraklion, Crete, Greece. 2 University of Athens, Department of Physics, Division of Applied Physics, Panepistimioupolis, Build. PHYS-V, GR-15784, Athens, Greece. ABSTRACT Digital Elevation Models (DEMs) and land cover products are primary inputs for hydrologic models of surface runoff that affects infiltration, erosion, and evapotranspiration. DEM and land cover play important role in determining the runoff characteristics of specific catchment areas. Recently, at local level, a number of data sources have been used to derive land cover products for high resolution studies. These studies have been carried out for a number of different applications, including estimation of biomass and vegetation mapping. A hydrologic land cover classification includes information not only about vegetation species, but also about the land surface and what classes are important hydrologically. This kind of classification must therefore incorporate information on elevation, slope, aspect, surface roughness, as well as vegetation species derived from satellite added-value products. The main problems when generating hydrologic land cover maps is the lack of accurate DEMs and the confusion of spectral responses from different features. In this study, a Terra/ASTER image acquired over the region of Heraklion, Crete, Greece was used. ASTER stereo imagery is used for DEM production because it gives a strong advantage in terms of radiometric variations versus the multi-date stereo-data acquisition with across-track stereo, which can then compensate for the weaker stereo geometry. GCPs (Ground Control Points) derived from differential GPS measurements were also used for absolute DEM production. A hydrologic land cover classification scheme was developed by combining ASTER multispectral imagery, ASTER DEM products and the spectral signatures derived from field observations at predefined training sites. Keywords: Digital Elevation Model, GPS, Multispectral Satellite Imagery, Hydrologic Land Cover Classification 1. INTRODUCTION Remote sensing provides a means of observing hydrologic state variables over large areas. Drainage basins, catchments and sub-catchments are the fundamental units for efficient management of land and water resources. Catchments and watersheds have been identified as planning units for administrative purposes to conserve these precious resources 1. The concept of watershed management recognises the inter-relationships between land cover/use, soil and water and the link between uplands and downstream areas. Watersheds are subject to hydrologic processes affected by the spatial variability of many parameters such as soils, topography, land use and cover, climate and human-induced changes and management. Landscape roughness refers to the unevenness of the earth s surface due to natural process (e.g. topography, vegetation, erosion) or human activities (e.g. buildings, power lines, forest clearings). Roughness affects transport of hydrometeorological fluxes between land surface and the atmosphere, as well as below the surface, i.e. infiltration and water movement. * zedd2@iacm.forth.gr; phone +30 2810 391762; fax +30 2810 391762; www.iacm.forth.gr/regional

Roughness is often classified in different complexities related to its effects on land surface atmosphere dynamics. These complexities are: a) vegetation and urban roughness where the horizontal scale is relatively small, b) transition roughness between landscape patches and c) topographic roughness due to changing landscape elevations 2. Digital Elevation Models (DEMs) are among the remote sensing techniques that have been used to measure landscape surface roughness properties over large areas. DEMs are increasingly used for visual and mathematical analysis of topography, landscapes and landforms, as well as modeling of surface processes 3, 4, 5. DEMs of usable detail are still not available for most of the Earth s surface, and when available, they usually lack sufficient accuracy. For the production of DEMs from optical satellite data, the respective satellite sensors should have stereo coverage capabilities. To obtain stereoscopy with images from satellite scanners, two solutions are possible 6 : a) the along-track stereoscopy with images from the same orbit using fore and aft images; and b) the across-track stereoscopy from two different orbits. The latter solution was the most widely used since 1980: first, with Landsat from two adjacent orbits, then with SPOT using across-track steering capabilities, and finally with IRS-1C/D. In the last few years, the former solution was used with the JERS-1 OPS, the German Modular Opto-Electronic Multi-Spectral Stereo Scanner (MOMS) and the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER). The viability of stereo correlation for parallax difference/height measurement from digital stereoscopic data has been described and evaluated in previous studies 7, 8, 9. AI-Rousan and Petrie 10 evaluated commercial software, for the extraction of DEMs using SPOT data. Cheng, et al. 11 generated a DEM from raw IRS-1C LISS stereo-images, over a mountainous area in Arizona. Toutin and Cheng 12 presented results for extraction of DEM from very high spatial resolution IKONOS images, and high spatial resolution ASTER and Landsat-7 ETM+ images. Commonly used statistical measures of map accuracy are the Root Mean Square Error (RMSE) in xy (planimetry) and z (elevation). As pointed out by Ackermann 8 and Bolstad and Stowe 13, RMSExyz alone is an insufficient measure of DEM data quality for those who use the DEM data for scientific applications. Other parameters include: a) The spacing of the DEM grid in xy (also referred to as posting, pixel size or xy precision). b) The resolution of the DEM in z (also referred to as z precision, and numerically the smallest z increment). c) Slope measurement accuracy (a derived value, dependent on RMSEz as well as the measurement distance in xy). d) Equivalent map scale based on map accuracy standards. Based on expected RMSEz values for ASTER DEMs, slope accuracies of 5 should be obtainable over measurement distances in the 100 m to 500 m and longer range. The production of land cover / land use maps is a common application of multispectral satellite images. There are numerous examples of land cover maps derived from multispectral satellite imagery at global, regional and local level. The most widely used approach is to classify each image pixel as an independent observation, regardsless of its spatial context. Recently, at local level, a number of data sources have been used to derive land cover products, including Landsat TM data for high resolution studies 14. These studies have been carried out for a number of different applications, including estimation of biomass 14 and vegetation mapping 15. Vani and Sanjeevi 16 evaluated the potential of ASTER VNIR (Visible and Near Infrared) and SWIR (Short Wave Infrared) sensors for land cover mapping. Information in the VNIR image contributed to the enhancement of vegetation and water classes. Rock and soil units were enhanced due to the contribution by the information in the SWIR images. One of the main problems when generating land cover maps from digital images is the confusion of spectral responses from different features. Sometimes two or more different features with similar spectral behaviour are grouped into the same class, which leads to errors in the final map. The accuracy of the map depends on the spatial and spectral resolution and the seasonal variability in vegetation cover types and soil moisture conditions 17. Attempts have been made to improve the accuracy of image classification. One approach is the use of multi-temporal imagery to distinguish classes 18, 19. Another one is the piecewise linear classifier with simple post-processing 20. The integration of GIS with ancillary information has also been tested, to improve image classification. Gastellu-Etchegorry et al. 21 and Ortiz et al. 22 integrated GIS with information about soils, topography and bioclimate. Similarly, Palacio-Prieto and Luna-González 23 employed GIS rules with ancillary data on terrain mapping units and elevation data.

A hydrologic land cover classification must include information not only on vegetation species, but also on land surface and what classes are important hydrologically. This kind of classification must incorporate information on elevation, slope, aspect, surface roughness, and vegetation species derived from satellite added-value products. In this study, high spatial resolution multispectral imagery from ASTER onboard Terra satellite was analyzed in combination with Global Positioning System (GPS) data and field observations to provide accurate DEM and land cover thematic maps and finally to develop a hydrologic land cover classification scheme for the area of NW Heraklion Prefecture in the island of Crete. The classification output is a thematic map of the study area presenting fifteen levels of surface runoff potential, which have been estimated with the combined use of the spectral information (NVIR, SWIR channels), as well as spatial and topographic information (DEM, slopes, aspects). 2. DATA AND METHODOLOGY The data used in this study were a Terra/ASTER image of August 10, 2000 at 09.25 LST, over the region of Heraklion, Crete, Greece, and GCPs (Ground Control Points) derived from GPS measurements from the differential GPS ground station of the Regional Analysis Division of FORTH/IACM (Foundation for Research and Technology Hellas, Institute of Applied and Computational Mathematics). The ASTER method is used because it gives a strong advantage in terms of radiometric variations versus the multi-date stereo-data acquisition with across-track stereo, which can then compensate for the weaker stereo geometry. Field observations have been also used for training site selection for the supervised classification of ASTER multispectral imagery. A Transverse Mercator projection was applied (Projection System: Hellenic Geodetic Reference System 87 - HGRS87) in order to have all data in the same cartographic projection system. ASTER is an advanced multispectral imager that was launched on board NASA s Terra spacecraft in December, 1999. ASTER covers a wide spectral region with 14 bands from the visible to the thermal infrared with high spatial, spectral and radiometric resolution. ASTER consists of three separate subsystems, each operating in a different spectral region, using separate optical systems. These subsystems are the VNIR, the SWIR and the thermal infrared (TIR). The spatial resolution varies with wavelength: 15 m in the VNIR, 30 m in the SWIR and 90 m in the TIR. The VNIR subsystem consists of two telescopes one nadir looking with a three band detector (Channels 1, 2 and 3N) and the other backward looking (27.7 off-nadir) with a single band detector (Channel 3B). The SWIR subsystem consists of one telescope with a six band detector (Channels 4 to 9). The TIR subsystem consists of one telescope with a five band detector (Channels 10 to 14). The most important specifications of the ASTER stereo subsystem that govern the DEM generation capabilities, include: stereo geometry; platform altitude of 705 km and ground speed of 6.7 km/sec; B/H (base-to height ratio) of 0.6; IFOV of 15 m; bandpass of 0.76-0.86 micrometers in channel 3N and 3B, both with an MTF of 0.24; 9 seconds required to acquire a 60 x 60 km scene; 64 seconds required to acquire a stereo pair 7. The nadir-backward stereo-viewing geometry potentially gives a higher probability of obtaining a cloud free image pair, as compared to a side stereo observation system requiring multi-orbit observation, such as SPOT. DEM extraction from ASTER stereo imagery is based on the principle of automatic stereo correlation. The accuracy to which absolute elevations can be derived by photogrammetric techniques is governed by: a) B/H ratio i.e. geometric stereo disposition, b) reliability of the correlation procedure and c) accuracy and density of ground control points. DEM accuracy depends mainly on the following factors 24 : sensor model, image deformations, geometric accuracy and resolution of scanner as well as accuracy of interior orientation. Furthermore, the number, accuracy and regular horizontal and vertical distribution of GCPs are important. ASTER 3N and 3B images were used for DEM production, whereas VNIR and SWIR imagery were used for land cover classification and mapping. Field measurements with the use of differential GPS were also performed to provide GCPs for DEM correction and geo-location, as well as to support field observations and training site selection necessary for supervised classification. A digital stereo correlation approach was used to calculate parallax differences from ASTER stereo pair. The mathematical concept of one approach to stereo correlation is given by Lang and Welch 7. Relative ground elevations were determined by measuring the parallax differences in the registered images. The parallax differences were converted to absolute elevations with the use of GCPs. The image coordinates of the GCPs in conjunction with their HGRS87 map coordinates allowed the development of transformation equations needed to register the stereo images and eventually geodetically rectify them to the Earth's surface. The geometric model used was a rigorous parametric model developed at the Canada Centre for Remote Sensing 25. Assuming parallax difference correlation errors in the range of 0.5 to 1.0 pixels (7-15 m), elevation errors (RMSEz) would be in the ±12 m to ±26 m

range. The Root Mean Square Error (RMSE) in XY (planimetry) and Z (elevation) were used for the DEM accuracy assessment. Slope and aspect images for the area of concern were produced from the DEM using ERDAS Imagine commercial software. The slope image illustrates changes in elevation over a certain distance. This distance is the size of the pixel (15 m). Slope is most often expressed as a percentage, but can also be calculated in degrees. For a pixel at a given location, the elevations around it were used to calculate the slope. In practice, 3x3 pixel window (45 m x 45 m) was used to calculate the slope at each pixel. The aspect image illustrates the prevailing direction that the slope faces at each pixel. Aspect is expressed in degrees from north, clockwise, from 0 to 360. Due north is 0 degrees. A value of 90 degrees is due east, 180 degrees is due south and 270 degrees is due west. Consequently, a value of 361 degrees is used to identify flat surfaces such as water bodies. As with slope calculations, aspect uses a 3 x 3 window around each pixel to calculate the prevailing direction it faces. Figure 1: The hydrologic land cover classification scheme.

Following, a supervised classification was applied to ASTER VNIR and SWIR imagery to generate a land cover map. The differential GPS was used for the positioning of training areas and for the collection of ground truth data during a field observation campaign. Ten land cover classes were initially identified: a) Water bodies, b) urban areas, c) sclerophylous vegetation, d) natural grassland, e) olive trees, f) vineyards, g) complex cultivation patterns and three classes for bare soil. The Bayesian Maximum Likelihood Classifier was used in this classification process. Next a post classification sorting took place and seven input classes for the hydrologic land cover classification scheme were produced: a) Bare soil, b) grass, c) vineyards, d) complex cultivation patterns, trees, urban areas and water bodies. These classes are hydrologically important, because each one represents a different type of surface roughness. Finally, a classification scheme providing a hydrologic land cover categorization of the study area was developed. A Knowledge Base was developed by combining the supervised classification result (the aforementioned seven classes) with the topographic information (elevation, slope and aspect) to estimate the surface runoff potential. The runoff potential is dependent on surface type (i.e. it is higher for bare soil than for trees). For a given surface type, the higher the slope the higher the runoff potential. In practice, slopes were grouped in three main categories: a) higher than 15 degrees, b) between 5 and 15 degrees and c) between 0 and 5 degrees. For a given surface type and slope, the runoff potential is higher for those surfaces facing the main streams. Based on the aspect information for each pixel, all surfaces were grouped in two main categories: a) the perpendicular ( ) surfaces and b) the parallel ( ) surfaces. Imagine an axis perpendicular to the plane of each surface. Let φ be the angle of the projection of this axis to the plane of the nearest stream with the axis of this stream. If φ is between 0 and 45 degrees or between 135 and 180 degrees, the respective surface is grouped as perpendicular. If φ is between 45 and 135 degrees, the respective surface is grouped as parallel. After these assumptions for slope and aspect, 15 runoff potential classes (level 1 to level 15) were identified in the final classification scheme as shown in Figure 1. 3. RESULTS ASTER NIR stereo imagery was used for DEM production. Stereo GCPs were available within ± 2 m accuracy. PCI photogrammetric software (Orthoengine) was employed for DEM production. Figure 2 shows the produced DEM for the NW Heraklion area. Catchments and mountainous areas are clearly depicted. DEM planimetric accuracy was estimated at ± 15 m (1 ASTER pixel), whereas its vertical accuracy was estimated using 200 trigonometric points, available from 1:5000 topographic maps for the prefecture of Heraklion. The elevation of each trigonometric point was compared with the elevation of the respective DEM pixel and it was founded that (with 95 % confidence) the RMSE was ± 11 m. GIS techniques were used for slope extraction from the produced DEM. The planimetric accuracy of the produced DEM was at ± 15 m (1 ASTER pixel), whereas its vertical accuracy was checked by using 100 trigonometric points, available from 1:5000 topographic maps for the prefecture of Heraklion. The RMSE (Root Mean Square Error) was used for the evaluation of DEM vertical accuracy. RMSE encompasses both systematic and random errors and is defined as: RMSE = 1 n n = i 1 δ 2 z i where, δz i n is the elevation differences between the derived DEM and the trigonometric points is the number of trigonometric points. The elevation of each trigonometric point was compared with the elevation of the respective DEM pixel and it was founded (with 95 % confidence) that the RMSE was 9.52 m, whereas the maximum error was 18 m and the mean error was 8.20 m. From the statistical results it was clear that: a) The majority (80%) of the differences between the computed elevation values and the input elevation values are less than 15 m.

b) The mean error is less than 8.5 m, showing that the major differences are a few meters, larger errors are the minority. c) The maximum error is 18 m, an acceptable value yielding that the distribution of errors is within the normal extent. Figure 2: DEM for the NW Heraklion area as has been derived from ASTER stereo imagery. Slopes were estimated from the DEM for the study-area. Figure 3 shows the slope image of the area. Slope images are usually color-coded according to the steepness of the terrain at each pixel. In Figure 3, the calculated slopes have been grouped in seven main classes. Using the elevation and the slope information, aspect values for each pixel were calculated and all land surfaces of the study area were classified in two main categories as described in section 2. In Figure 4, the surfaces with axes perpendicular to main streams are presented in red, whereas the surfaces with axes parallel to the main streams are presented in green.

Figure 3: Slope image of the study area as has been produced from DEM analysis. Figure 4: Surface orientation in relation to the nearest main streams.

Figure 5: The final land cover product for the NW Heraklion area. Spectral signatures derived from ASTER multispectral imagery, as well as elevation, slope and aspect information derived from ASTER stereo imagery have been used. The final classification product is presented in Figure 5. Urban and water classes have been produced from the initial supervised classification. The final classification scheme was not applied to the urban surfaces due to the sub-pixel variations of slope and aspect in urban environment. These variations are significant error sources when the spatial resolution of the used sensors is of the order of magnitude of 15 m. As far as the estimated runoff classes are concerned, the percentage area covered by each runoff class is presented in Table 1. It should be noted that the 50% of the total area corresponds to the first 4 runoff classes. As a result the runoff potential is rather low for the study area as a whole, but there are a lot of localized areas with high runoff potential. The detailed mapping of these areas is a valuable information for the parametrization of hydrologic models dealing with surface processes. Table 1. Percentage area covered by each class Runoff Level Area (%) 1 13.78 2 13.30 3 9.75 4 12.29 5 15.82 6 8.33 7 5.08 8 5.76 9 3.34 10 5.01 11 3.09 12 4.02 13 0.16 14 0.20 15 0.01

4. CONCLUSIONS In this study, high spatial resolution ASTER imagery was analyzed in combination with GPS data and field observations to provide a hydrologic land cover classification capable of supporting hydrologic modeling. ASTER data were used because it gives a strong advantage in terms of radiometric variations versus the multi-date stereo-data acquisition with across-track stereo, which can then compensate for the weaker stereo geometry. Near infrared stereo imagery was used for DEM production, whereas VNIR and SWIR imagery were used for land cover classification and mapping. Field measurements were carried out to provide GCPs for DEM correction and geo-location, as well as to support field observations and training site selection for the necessary for supervised classification. Likewise, a hydrologic land cover classification scheme was applied by combining the vegetation classes provided by the supervised classification process with the slope and aspect values for each pixel provided by the DEM. The result of the DEM evaluation showed an acceptable accuracy for runoff potential levels categorization. The planimetric and elevation accuracy of the produced DEM (± 15 and ± 9.52 m, respectively) are considered quite satisfactory for large catchment hydrologic parameterization, indicating the high potential of ASTER imagery to support hydrologic modeling. REFERENCES 1. M. A. Khan, Water balance and hydrochemistry of precipitation components in forested ecosystem in the arid zone of Rajasthan, India, Hydrological Science Journal, 44, pp. 149 161, 1999. 2. T. Schmugge, A. French, J. C. Ritchie, A. Rango and H. Pelgrum, Temperature and emissivity separation from multispectral thermal infrared observations, Remote Sens. Environ., 79, pp. 189-198, 2002. 3. M. P. Bishop and J. F. Shroder, Remote sensing and geomorphometric assessment of topographic complexity and erosion dynamics in the Nanga Parbat massif, In: Tectonics of the Nanga Parbat Syntaxis and the Western Himalaya (M. A. Khan, P. J. Treloar, M. P. Searle and M. Q. Jan, eds.), Geological Society London, Special Publications, 170, London, pp. 181-199, 2000. 4. G. C. Millaresis and D. P. Argialas, Extraction and delineation of alluvial fans from digital elevation models and Landsat Thematic Mapper images Photogrammetric Eng. Remote Sens., 66, pp. 1093-1101, 2000. 5. G. E. Tucker, F. Catani, A. Rinaldo and R. L. Bras, Statistical analysis of drainage density from digital terrain data, Geomorphology, 36, pp. 187-202, 2001. 6. Th. Toutin Th., Elevation Modelling from Satellite VIR Data: A Review, Int. J. Remote Sens., 22, pp. 1097-1125, 2001. 7. H. Lang and R. Welch, R., ATBD-AST-08 Algorithm Theoretical Basis Document for ASTER Digital Elevation Models, Standard Product AST14 Report, The Jet Propulsion Laboratory, California Institute of Technology, Los Angeles, USA, 1999. 8. F. Ackermann, F., Digital elevation models techniques and application, quality standards, development, Int. Arch. Photogrammetry Remote Sens., 30, pp. 421-432, 1994. 9. P.V. Bolstad and T. Stowe, An evaluation of DEM accuracy: Elevation, slope, and aspect Photogramm. Eng. Remote Sens., 60, pp. 1327-1332, 1994. 10. N. AI-Rousan and G. Petrie, System calibration, geometric accuracy testing and validation of DEM and orthoimage data extracted from SPOT stereopairs using commercially available image processing systems, Int. Arch. Photogrammetry Remote Sens., 32, pp. 8-15, 1998. 11. Ph. Cheng, Th. Toutin and C. Stohr, Automated DEM Extraction from Aerial Photos or Optical Satellite Images, Proc. of the 13 th International Conference on Applied Geologic Remote Sensing, March 1-3, 1999, Vancouver, Vol. 1, pp. 56-63, 1999. 12. Th. Toutin and Ph. Cheng, Ph., Demystification of IKONOS, EOM Archives, 2000, URL: http://www.eomonline. com/common/archives/ July00/toutin.htm. 13. P. V. Bolstad and T. Stowe, An evaluation of DEM accuracy: Elevation, slope, and aspect Photogrammetric Eng. Remote Sens., 60, pp. 1327-1332, 1994. 14. M. K. Steininger, Satellite estimation of tropical secondary forest above-ground biomass: data from Brazil and Bolivia, Int. J. Remote Sens., 21, pp. 1139 1175, 2000. 15. D. Muchoney, J. Borak, H. Chi, M. Friedl, S. Gopal, J. Hodges, N. Morrow and A. Strahler, Application of the MODIS global supervised classification model for vegetation and land cover mapping of Central America, Int. J. Remote Sens., 21, pp. 1115 1138, 2000.

16. K. Vani and S. Sanjeevi Fusion of ASTER image data for enhanved mapping of landcover features, The Geographic Information Systems Portal, 2002, URL: http://www.gisdevelopment.net/application/environment/ pp/envp0005.htm. 17. R.G. Congalton, Using spatial autocorrelation analysis to explore the errors in maps generated from remotely sensed data, Photogramm. Eng. Remote Sens., 54, pp. 587-592, 1998. 18. C. Conese and F. Maselli, Use of multi-temporal information to improve classification performance of TM scenes in complex terrain, ISPRS J. Photogramm. Remote Sens., 46, pp. 187-197, 1991. 19. P. M. Barbosa, M. A. Casterad and J. Herrero, Performance of several Landsat 5 Thematic Mapper (TM) image classification methods for crop extent estimates in an irrigation district, Int. J. Remote Sens., 17, pp. 3665-3674, 1996. 20. K. Y. Huang and P. Mausel, Spatial post-processing of spectrally classified video images by a piece wise liner classifier, Int. J. Remote Sens., 14, pp. 2563-2573, 1993. 21. J. P. Gastellu-Etchegorry, C. Estreguil and E. Mougin, A GIS based methodology for small-scale monitoring of tropical forests - A case study in Sumatra, Int. J. Remote Sens., 14, pp. 2349-2368, 1993. 22. M. J. Ortiz, R. Formaggio and J. C. N. Epiphanio, Classification of croplands through integration of remote sensing, GIS and historical database, Int. J. Remote Sens., 18, pp. 95-105, 1997. 23. J. L. Palacio-Prieto and L. Luna-González, Improving spectral results in a GIS context, Int. J. Remote Sens., 17, pp. 2201-2209, 1996. 24. A. Altmaier and C. Kany, Digital surface model generation from CORONA satellite images, ISPRS J. Photogramm. Remote Sens., 56, pp. 221-235, 2002. 25. Th. Toutin, Multi-source Data Fusion with an integrated and Unified Geometric Modelling, Advances in Remote Sensing EARSeL Journal, 4, pp. 118 129, 1955.