Improving Land-cover Classification with a Knowledge- Based Approach and Ancillary Data

The classification of land cover is one of the major applications of remotely sensed data. However, the distinction of different classes is often challenging because of spectral similarities. Numerous studies are published describing advanced methods for separation of spectral similar land-cover classes from the remote sensed images. But there are still limitations in terms of classification ac-curacy when using only spectral information. In this paper a method is described which improves classification results by far using additional information. In detail global available digital elevation data are utilized as an additional information layer to separate between spectral very similar classes which are yet ecological very different. The base idea is that the vegetation types are closely related to the ecological conditions. These ecologic conditions are inter alia a function of the topography. By processing the raw elevation data the topography at every point can be described by land form classes (e.g. hill, slope, depression et al.). The so gained additional information source is used to refine the result of a maximum likelihood classification as well as an additional layer during the classification process. With this approach the classification accuracies for the problematic classes can be improved up to nearly 50%. INTRODUCTION Accurate land-cover products are very important inputs for the assessment of the estate of the landscape as well as for environmental models describing erosion, hydrological, climate or biodiversity. Especially in countries with no detailed available land-cover statistics, the derivation of land-cover products for large areas on the basis of remote sensing data is a highly important task (i). Such data sets are received through classification of multispectral satellite data. One of the main problems during classification arises because the target objects (information classes) are not explicit spectrally separable, which is a prerequisite for multispectral image classification. Reasons can be inadequate spatial resolution of the sensor resulting in mixed pixels or spectral very similar characteristics of different land-cover types. Many different classification algorithms exist to separate classes with small spectral differences e.g. spectral mixture modelling (ii ,iii). These approaches are working with pure spectral characteristics of the desired land-cover types which are not available in our study area. Further more, if the land-cover types are spectral similar but internal inhomogeneous results of traditional classifier are typically intermixed by the similar classes. One solution to the problem is the incorporation of ancillary data to improve the classification. That approach relies on the ecological fact, that the occurrence of specific vegetation types are often bounded to particular environmental parameters such as elevation, aspect, soil etc. By incorporating such additional information which determines the ecological livelihoods of the vegetation units, classification accuracy can be improved. Most of the works that followed this approach used environmental data like elevation (iv ,v), soil characteristics (vi) or other ecological parameters (vii). All of them used the ancillary information after the proper multispectral classification to refine broader land-cover classes. In this work topographic informaCenter for Remote Sensing of Land Surfaces, Bonn, 28-30 September 2006 185 tion is included during the classification process and results are compared to postclassification improvement results. STUDY AREA AND METHODS To test the benefit of incorporating ancillary data into the classification process, remote sensing data from central Benin (Landsat ETM+, 26.10.2000) as well as digital elevation data (SRTM data, 3 arc-seconds resolution, approx. 91m2) are used. The two study areas are located in the sub-humid tropics, showing a typical undulating relief (pediplain) with very low elevation differences. The quasi-natural vegetation is dominated by a forest-savannah mosaic with azonal vegetation types along water courses (gallery forests) and on inselbergs (viii). The region is subject to heavy agricultural colonization due to high population growth rates (ix). During a supervised land-cover classification (x) it was found, that some land-cover types are spectrally very similar and not easy separable. Especially the distinguishing between gallery forest and non-gallery forest are often not possible (Fig. 1) as well the discrimination between inselbergs and some agricultural field areas is difficult. Both land-cover types posses special eco-logical attributes and behavior and should be distinguished in the land-cover classification. Gallery forest and inselbergs are restricted to particular locations in the landscape, namely valleys and isolated (rocky) mountains respectively. Those topographic features cannot be identified solely by pure elevation and / or slope data, e.g. flat terrain is occur-ring on the valley bottom as well as on top of hills. So, more integrative landscape form descriptions are necessary. There exists many approaches to calculate landscape form parameters based on digital elevation models in geomorphology, but often they are more or less complex as they incorporate several convex / concave descriptors and deliver numerous classes (xi ,xii ,xiii). As the vegetation types in the study area are not dependent from such detailed landforms a topographic position index (TPI) is calculated (xiv ,xv). This index calculates the relative position of every pixel based on digital elevation data. The relative position is defined as the difference of the height value of a location from the surrounding mean height values. The size of the neighborhood which is incorporated in the calculation is de-pending of the desired detail of the index. The calculation can be done with any focal window shape. The resulting index values are negative or positive with a range of value depending on the height differences in the focal window. Negative values indicate that the local position is deeper, positive values indicate that the location is higher than the neighborhood. Zero values occur at flat plains or at mid-slope positions. Based on that index values a very simple land from classification is possible, e.g. valley, slope and ridge. The results of different sizes of focal windows can be used to derive more complex landform classes’ e.g. small valley on a flat hill (xv). To facilitate the comparison between several focal sizes the raw index values are standardized by subtracting the mean and dividing by standard deviation. Calculation of the topographic position index and classification was done with the Bentic Terrain Modeller extension for ArcGIS (xvi), which is designed for subaqueous landscapes but is also applicable to terrestrial landscapes. Fig. 1: spectral curve for gallery forest and dry forest. Min and max ranges are shown as vertical lines. Proceedings of the 2 Workshop of the EARSeL SIG on Land Use and Land Cover 186 In this study the two focal window sizes of 10 and 30 pixels are used to derive TPI data. The results were used to derive a landform classification of the study areas with seven classes. The used thresholds were derived from empirical analysis of the area and are shown in Tab. 1. There are two possibilities to incorporate the topographic information into the classification process. A land form classification product can be used to split mixed multispectral land-cover classes in a post-classification knowledge based approach, or the standardized index values can be used as additional information source for the training of the classifier (multi-layer approach). The land-cover classification was performed with a supervised maximum likelihood (ML) classification. Both methods are performed and evaluated against classifications without TPI information with independent reference areas in an error matrix. In each case the classification was performed with the same set of training areas to ensure comparability. To perform the method two subsets from a Landsat scene (192-54) were created each with typical representation of gallery forest (study area 1) and inselbergs (study area 2) respectively. RESULTS & DISCUSSION Improvements of class gallery forest The ML classification of study area 1 shows clearly the mixture of the classes “gallery forest” and “dry forest” (Fig. 3, a1). Especially dry forest islands alongside of rivers are often classiFig. 2: Schematic illustration of resulting TPI values according to relief. Tab. 1: Thresholds to derive landform classes from TPI data. Values are standard deviations*100. Center for Remote Sensing of Land Surfaces, Bonn, 28-30 September 2006 187 fied as gallery forest (see focus 1 in Fig. 3) which is obviously false regarding field data and digital elevation values. But there is also the contrary that gallery forest is classified as dry forest (focus 2 in Fig. 3). Accuracy values are given in Tab. 2. Classification was performed with some other land-cover classes to deliver a holistic view of the area and to “feed” the classifier with training data for the whole range of data of the multispectral image. The first experiment to eliminate the false classified pixels uses the landform classes derived from TPI data. The ML result was refined by simple conditional queries so that dry forest pixels inside valleys and gallery forest pixels outside valleys are changed (Fig. 3, a2). The result gives much better accuracies (see Tab. 3), but is dependent from the landform classification. If a valley is not clearly classifiable from the TPI data, then dry forest will remain in the valley location (see focus 1) and will lead to worse results compared to the standard ML classification (in this case). But dry forest islands adjacent to gallery forests are clearly separable with this method. The second experiment used the standardized TPI data as additional layer as input for the ML classification. Results are much clearer regarding the shape of the gallery forests (Fig. 3 a3). This is due to the fact that low TPI values are used in combination with multispectral data as training data for the ML classification. If no multispectral signatures of gallery forests occur inside valleys, then also no gallery forest is classified. Accuracies exceed those of the former method (see Tab. 4). But the following problems are observed: If the shape of the valley is very flat, which is a common case in this region, the width of the gallery forests are sometimes to big especially in the case of continuous transition to dry forests. Improvement of class Inselberg The ML classification of pure multispectral data of study area 2 reveals comparable confusion between similar land-cover types (Fig. 3 b1). Some field areas are classified as inselberg whereas inselberg areas are sometimes classified as field or forest (in the case of vegetated inselberg, northern white arrow) (see Tab. 5). This misclassification leads to problems if those Proceedings of the 2 Workshop of the EARSeL SIG on Land Use and Land Cover 188 data are used to derive agricultural statistics. The same two experiments as above are carried out to get rid of those misclassified pixels. Fig. 3: Study areas printed in false color (4,3,2) and classification results (a). Result of maximum likelihood classification of multispec-tral data (b). Colours of land-cover classes: gallery forest = violet; inselbergs =blue; dry forest = dark green; savannah = light green,fields and others = brown. Gallery forest is merged with forest in study area 2. White arrows indicate inselbergs. Center for Remote Sensing of Land Surfaces, Bonn, 28-30 September 2006 189 The first improvement for study area 2 was the post-classification refinement with land form classes. Results are very clear, but inselberg areas tend to be overestimated. However, false positives are eliminated to a high fraction (Fig. 3, b2). Smaller inselbergs are not identified by that method as well the special case that the inselberg has a rocky surface but not a clear morphological shape (SW white arrow). The incorporation of standardized TPI data into the ML classification process produced the best results in form of accuracy and spatial representation of inselbergs (Fig. 3 b3; Tab. 7). Smaller inselbergs are better included but at the cost of some false positives in agricultural area. But also the difficult case that rock outcrops appear in the middle of field is captured with this method (compare middle white arrow). Naturally, if the surface of field areas and inselbergs are identical but no morphological difference exists, the method will fail. This is the case between in the area between the middle and south western arrow. The better representation of gallery forest in this study area is a comfortable side effect of the additional TPI data in the ML classification. CONCLUSIONS The use of additional environmental and neighborhood information during a maximum likelihood classification process or in a post-classification refinement showed a tremendous improvement of land-cover classification accuracy of up to 50 % for intermixed classes. The two methods explored are very different and have their own limitations. If predefined land form classes are used, then the results will highly depend on those classification definitions. If landforms are not clearly separable, e.g. if the height of gallery forest are visible in the digital elevation model (what is sometimes the case even in the SRTM data), appropriate class assignments are difficult. The use of topographic position index values as additional layer during maximum likelihood classification result in the best separation of spectral similar classes. However, there is one general problem with this method: if one class has no correlation to topography, i.e. the class has no correlation to a specific position in the landscape, a training set for every of those locations are necessary to train the classifier appropriately. This obserProceedings of the 2 Workshop of the EARSeL SIG on Land Use and Land Cover 190 vation was already reported (iv). But the presented approach is nevertheless very useful to separate spectral intermixed classes and it should be possible to apply the method in other regions worldwide, as SRTM data are available globally. ACKNOWLEDGEMENTSThis work was supported by the Federal German Ministry of Education and Research (BMBF)under grant No. 01 LW 0301A and by the Ministry of Science and Research (MWF) of thefederal State of Northrhine-Westfalia under grant No. 223-21200200. REFERENCESi Lautenbacher C, 2006. The Global Earth Observation System of Systems: Science Serv-ing Soci-ety, Space Policy 22(1), 8-11. ii Adams J, D Sabol, V Kapos, R Filho, D Roberts, M Smith, & A Gillespie, 1995.Classification of multispectral images based on fractions of endmembers: Application toland-cover change in the Brazilian Amazon, Remote Sensing of Environment 52(2), 137-154. iii Kuemmerle, T, A Roder, & J Hill, 2006. Separating grassland and shrub vegetation bymultidate pixel-adaptive spectral mixture analysis, International Journal of RemoteSensing 27(15), 3251-3271. iv Fleming M & R Hoffer, 1979. Machine Processing Of Landsat MSS Data And DMATopographic Data For Forest Cover Type Mapping, in: Proceedings of the 1979 MachineProcessing Of Re-motely Sensed Data Symposium, pp. 377—390. v Stolz R, M Braun, M Probeck, R Weidinger, & W Mauser, 2005. Land use classification incom-plex terrain: the role of ancillary knowledge, EARSeL eProceedings 4(1), 94-106. vi Baker C, R Lawrence, C Montague, & D Patten, 2006. Mapping wetlands and riparianareas using Landsat ETM+ imagery and decision-tree-based models Wetlands , 26, 465-474 vii Loveland T R, J W Merchant, D O Ohlen, & J F Brown, 1991. Development of a land-cover characteristics database for the conterminous united-states, PhotogrammetricEngineering And Remote Sensing 57(11), 1453-1463. viii Orthmann B, 2005. Vegetation ecology of woodland-savanna mosaic in central Benin(West Africa): Ecosystem analysis with focus on the impact of selective logging, PhDthesis, University of Bonn. ix Thamm H-P, M Judex, & G Menz, 2005. Modeling of Land-Use and Land-Cover Change(LUCC) in Western Africa using Remote Sensing, Photogrammetrie FernerkundungGeoinformation 3/2005, 191-199. x Thamm H-P, G Menz & O Kissiyar, 2001. Investigation of the Land Use / Land CoverChange in the Upper Ouémé Catchment, Benin (West-Africa) for the Set Up of aCoherent Development Plan. In: Proceedings of the 2001 IGRASS Symposium, Sydney xi Dikau & Schmidt 1999 xii Schmidt J & R Dikau, 1999. Extracting geomorphometric attributes and objects from digi-tal ele-vation models – semantics, methods, future needs. In: GIS for earth surfacesystems – analysis and modelling of the natural environment edited by Dikau R & HSaurer. (E. Schweizbart’sche Ver-lagsbuchhandlung, Stuttgart) 153–73 xiii Wilson J & J Gallant, 2000. Terrain analysis: principles and applications Wiley,Chichester Center for Remote Sensing of Land Surfaces, Bonn, 28-30 September 2006 191xiv Weiss A, 2001. Topographic Position and Landforms Analysis. Poster presentation, ESRIUser Conference, San Diego, CA. xv Lundblad E R, D J Wright, J Miller, E M Larkin, R Rinehart, D F Naar, B T Donahue, S MAnder-son, & T Battista, 2004. Classifying Deep Water Benthic Habitats around Tutuila,American Samoa, in Proceedings of the Twenty-Fourth Annual ESRI User Conference. xvi Rinehart R W, D J Wright, E R Lundblad, E M Larkin, J Murphy & L Cary-Kothera, 2004.ArcGIS 8.x Benthic Terrain Modeler: Analysis in American Samoa, in Proceedings of theTwenty-Fourth Annual ESRI User Conference.