URBAN LAND COVER AND LAND USE CLASSIFICATION USING HIGH SPATIAL RESOLUTION IMAGES AND SPATIAL METRICS Ivan Lizarazo Universidad Distrital, Department of Cadastral Engineering, Bogota, Colombia; ilizarazo@udistrital.edu.co ABSTRACT In this paper, the author evaluates the results of applying two evolving classification techniques, decision trees (DT) and artificial neural networks back-propagation (ANN-BP), to obtain land cover and land use classes from Quickbird XS data combined with spatial metrics, in a selected urban zone in Bogota, Colombia. In order to improve the quality of the land cover classification, additional bands of texture and edges were added to the spectral bands and tested. In summary, results using the new methods did not surpass those obtained using the traditional maximum likelihood (ML) classifier. Nevertheless, the addition of one edge band rose the thematic accuracy of the classification compared to merely using the original spectral bands, no matter which classifier was used. The author also reports the attainment of good quality results when inferring urban land use from land cover units composition and diversity bands derived from the previous classification. In the test, the application of ANN-BP and DT algorithms did lead to get more accurate urban land use classes than using the ML classifier. Spatial metrics seems to be an appropriate framework to describe and better classify urban landscapes. Emergent classification techniques are very promising and need to be further investigated. INTRODUCTION The new generation of optical multispectral sensors like IKONOS, Quick- Bird and Orbview-3 provide raw data with a spatial resolution suitable for urban land cover and land use classification. But higher spatial detail does not mean higher spectral richness and some limitations arise to get accurate classes. On the classification of urban land cover a major problem is to deal with the spectral mixture of similar physical materials present in different land cover types (i). On the classification of urban land use the challenge is to find the adequate contextual data needed to infer classes that represent a functional concept the human activity that happens on the land- (ii). In the work reported here a solution to both challenges was proposed and tested by: (1) adding an edge image to Figure 1: Study zone as seen from QuickBird satellite (color composite RGB321). 292
Center for Remote Sensing of Land Surfaces, Bonn, 28-30 September 2006 the original image bands in order to separate mixed land cover classes, and 2) obtaining and using spatial metrics to infer land use types. Spatial metrics, also known as landscape metrics, is a methodology suitable for describing land use structures (iii, iv). In this study, two alternative supervised classification algorithms -back propagation (BP) and decision trees (DT) - were evaluated and compared with the traditional maximum likelihood (ML). Table 1: QuickBird XS image specifications (v). Band Wavelength (n m) Spatial Resolution (m) Blue 450 520 2.4 Green 520 600 2.4 Red 630 690 2.4 Near IR 760 900 2.4 METHODS The source data are four spectral bands composing a QuickBird-XS image spanning the visible and near-infrared wavelengths- taken in February 2005. In Table 1 the main characteristics of the image are indicated. The study area is located in the northwest of Bogota, the capital of Colombia, delimited between longitudes φ1=74o07'20.77''w to φ2=74o04'41.18''w and latitudes λ1=4o40'19.14''n to λ2=4o37'38.92''n, and covers about 2416 hectares. In Figure 1 a RGB321 colour composition of the image is shown. It is apparent the variety of land cover and land use units existing in the zone. In Table 2, the land cover classification scheme is shown. In Table 3, land use types can be seen. Both were defined based on the Anderson s classification schema (vi). In Table 2, the graphic samples show that some land cover classes have similar spectral signature and that additional data is needed to be able to differentiate between them. In Table 3, it is apparent that, in most cases, spectral information is useless to classify land use. Therefore, a spatial metrics approach is a suitable way to accomplish that task. In Figure 2, the workflow used in this work is depicted. This decomposition of the land use classification process in three stages has proven to be very useful in similar studies as reported in the recent literature (vii). Table 2: Land cover classification schema Code Description Sample 147 Asphalt road 148 Concrete road 181 Roof I (asphalt) 182 Roof II (asbestos) 183 Roof III (aluminiun) 184 Roof IV (clay tiles) 185 Roof V (glass fiber) 311 Grass 325 Bush 521 Water Body 731 Bare soil 293
Table 3: Land use classification schema Code Description Picture Sample 110 Residential single 115 Residential multiple 120 Commercial 130 Industrial 140 Transport Terminal 160 Institucional 171 Recreational open 174 Recreational roofed 190 Open space 294
Center for Remote Sensing of Land Surfaces, Bonn, 28-30 September 2006 STEP RESULT Digital classification of the multi-spectral image plus edge image Land cover classes Analysis of composition and spatial configuration metrics describing land use A set of suitable metrics Proportion and Diversity of land cover classes Discovery and application of land cover metrics based rules to infer land use categories Land use classes A set of rules Figure 2: Three-step workflow to obtain land-cover and land-use units. RESULTS & DISCUSSION Regarding the land cover classification, the neural network approach (ANN-BP) achieved a global thematic accuracy of 79% whilst decision trees (DT) was 72%. The performance of the maximum likelihood algorithm (ML) was higher than the emergent algorithms and accounted to 82%. In either case, the addition of a Sobel edge band increased thematic accuracy around 10%. The addition of a texture band derived from a panchromatic image did not improve the quality of the results. In Figure 3, a visual representation of a selected zone of the land cover classification using each approach is shown. It can be seen that the performance of the classifiers is similar. It also may be stated that thematic accuracy is not good enough and that some confusion remains in classes with similar spectral signature. On regards to the land use classification, ANN-BP algorithm global thematic accuracy was 84% while DT algorithm was 69% and ML algorithm only 42%, taking a training sample of about 10% of the image size. When the sample was reduced to 5% - a most common figure-, the accuracy decreased in a different amount for each classifier: it was 74% for DT, 65% for ANN-BP y 40% for ML. The spatial metrics selected to take into account the spatial pattern which characterizes land use structures were land cover percentage and land cover diversity. In Figures 4 y 5, the bar charts show the ability of some proportion (percentage) metrics to differentiate between different land use units. For space reasons only the grass percentage (Figure 4) and the bush percentage (Figure 5) are shown. It is apparent than some land use units have similar percentages of one specific land cover class and this means spatial confusion. Therefore, it is necessary to include the complete set of percentages in order to attempt the separation of the units. 295
(a) ML algorithm (b) ANN BP algorithm (c) DT algorithm (d) QuickBird-XS Figure 3: (a)(b) (c) Land cover classification using different algorithms, (d) true color composite. Figure 4: Percentage of grass in each land-use unit In Figure 6 (a), it is shown the urban land use classification attained using ANN-BP. In Figure 6(b) the sample training (5% of the total area) are delimited on the original image. According to the results attained in the land use classification, it may be stated that: (1) the spatial metrics selected in this study are capable to describe the land use to some degree but not completely, and (2) the performance of each classifier has a dramatic influence on the classification results. In regards to the latter statement, the decision tree algorithm achieved the highest thematic accuracy. It is worth to note that this algorithm produces a set the rules which is understandable by humans. The same cannot be said of the ANNs although their rules may be re-used too. 296
Center for Remote Sensing of Land Surfaces, Bonn, 28-30 September 2006 Figure 5: Percentage of bush in each land-use unit Figure 6: (a) ANN-BP land use classification, (b) Land use training samples CONCLUSIONS Reported figures show that the accuracy of the urban landscape classification depends on: (1) the algorithms ability to build decision rules capable to separate complex clusters, and (2) the addition of non spectral data. But they also suggest that the key data the cues that human interpreters find in every image (ii) are yet expecting to be extracted using digital means. In regards to that concern, the results of this study confirm previous investigations (ii, iii, iv, viii) and show that the spatial metrics framework is very promising and that a more comprehensive investigation of the whole set of indexes is needed to better clarify its ability for urban landscape classification in different socio-economical environments. In author s opinion next studies would benefit of including landscape metrics to classify both land cover and land use. In addition, as it has been noted before by experts (ii) it is worth to refine (or enhance) the digital means currently used in remote sensing to derive landscape metrics from the image data. 297
ACKNOWLEDGEMENTS The author acknowledges the support received from Universidad Distrital s Centre for Research (CIDC) that provided the image data and the software tools used in this study. REFERENCES i Mesev V, B Gorte, P A Longley, 2000. Classification Algorithms and their Application to Urban Remote Sensing, Chapter 5 in Donnay, J P., M J Barnsley, and P A Longley (eds.), Remote sensing and urban analysis, Taylor and Francis: London, 71-89. ii Barnsley M J,L Møller-Jensen, and S. L.Barr, 2000. Inferring Urban Land Use by Spatial and Structural Pattern Recognition, Chapter 7 in Donnay, J P., M J Barnsley, and P A Longley, (eds.), Remote sensing and urban analysis, Taylor and Francis: London, 115-144. iii Herold M, G Menz, 2001. Landscape Metric Signatures (LMS) to Improve Urban Land Use Information Derived from Remotely Sensed Data. In: Proceedings of the 20th EARSeL Symposium Remote Sensing in the 22st Century: A Decade of Trans-European Remote Sensing Cooperation. iv Herold M, J Scepan, K Clarke, 2002. The use of remote sensing and landscape metrics to describe structures and changes in urban land uses. Environment and Planning A, volume 34, pages 1443 1458. v Digital Globe, Product Overview. Available in http://www.digitalglobe.com. Last visit on July 30 de 2006. vi Anderson J R, 1976. A Land Use and Land Cover Classification System for Use with Remote Sensing Data. USGS Open Report. vii Donnay J P, M J Barnsley, P A Longley, 2000. Remote Sensing and Urban Analysis, Chapter 1 in Donnay, J-P., M J Barnsley, and P A Longley, P.A., (eds.), Remote sensing and urban analysis, Taylor and Francis: London, 3-18. viii Barr S L, M J Barnsley, A M Steel, 2004. Quantifying the separability of urban land use via a structural pattern recognition system, Environment and Planning B, 31, 397-418. 298