Intra-Urban Land Cover Classification in High Spatial Resolution Images using Object-Oriented Analysis: trends and challenges

Intra-Urban Land Cover Classification in High Spatial Resolution Images using Object-Oriented Analysis: trends and challenges Carolina Moutinho Duque de Pinho carolina@dpi.inpe.br

Introduction What is the importance in classifying land cover on such a detailed scale of urban areas? Impervious soil mapping surface run-off and flood studies in urban areas. Use this information for the analysis of urban microclimate. Studies on urban vegetation urban greening maps of town neighborhoods. Act as a initial stage for land use classification processes.

Introduction The object-oriented analysis is applied in l intra-urban land cover classification in in many cities of the world. Many of them have reached a very good thematic accuracy. But is necessary to point out some issues in these researches: They have been using a few number of class classification systems less complex thus the possibilities of errors decreases. Many of them were realized in well planning cities, European and American. In those cities there are few numbers of spatial patterns well defined. Often, the differences among the patterns inside the test area are not very big. Generally, they use a very small area of study, many times the area is restrict a couple of quarters. Thus, It has a lesser amount of problems with computer processing capacity.

Prupose This presentation is committed to show the shortcomings and alternatives in intra-urban land cover classification using high resolution images, specially in Brazilian cities, where the urban planning and management have not been able to control the urban sprawl.

Test Area SÃO JOSÉ DOS CAMPOS TEST AREA (12 km 2 ) SÃO PAULO STATE

Experiments Experiment I carried out for a complex intrasetting; urban classification scheme has been conceived and further applied to the whole study area, a highly complex and heterogeneous environment. Experiment II accomplished for a smaller intraarea. urban The goal was to evaluate the influence of urban occupation on the performance of land cover classification. Five quarters of Sao José dos Campos with different spatial patterns were selected.

Experiments Data: Merged Quickbird image of may, 2004 are used; Vector data of blocks to restrict the occurrence of built areas land cover classes; Vector data of quarters; IHS composition from natural color image. Software Envi 4.0 for pre- processing tasks E- Cognition 4.0 for object- oriented image analysis.

Experiment I - Segmentation Set of parameters to Keep shape? Set of parameters to Keep spectral information?

Difficulties in Experiment I It was difficult to choose a set of segmentation parameters that match to all of the spatial patterns in the test area. Thus I have chosen between keep the shape of the manmade objects from well-organized quarters or keep the spectral information from the manmade objects and natural objects from disorganized quarters. I have chosen the first option. Results: A very large number of objects (approximately 400.000)create problems with computer processing limitations: It has hindered the re-segmentation operations. It has calculated slowly the sample histograms The limits of created objects did not translate the shape of them. Thus, I could not use the shape attributes to do the semantic net.

Experiment - I The Semantic Net objects Vegetation Non-Vegetation 11 classes Trees Grass Shadow Non-Shadow Brighter objects (light types of concrete; some cars; metallic roofs Non- Brighter objects Coloreds Non-Coloreds Brown Light Red Medium Concrete Swimming pool Bluish Dark Bare Soil Dark Ceramics Light Ceramics Light Bare Soil Asphalt Dark objects Dark Concrete Metallic roofs Ceramics Asphalt Pavement Bare Soil Error Asphalt Real Dark Concrete

Exp. I - Classification Exatidão z Kappa 25,93 Global 0,580 0,537 26,70 25,93 0,607 0,580 0,567 0,537 Classes Brilhantes Cerâmica Solo Exposto Cobertura Metálica Concreto / Amianto Médio Médio Concreto / Amianto Escuro Asfalto Piscina Sombra Vegetação Arbórea Vegetação Rasteira Objetos não Classificados Brighter Objects Ceramics Bare Soil Metallic Roofs Medium concrete Dark Concrete Asphalt Swimming Poll Shadow Trees Grass Non-Classified Objects

Kappa per class 1,20 1,00 1,00 0,80 0,75 0,74 0,71 0,60 0,61 0,40 0,20 0,00 0,59 0,55 0,49 0,47 0,42 0,28 Bare Soil Dark Concrete Ceramics Shadow Medium concrete Trees Class Swimming Pool Grass Metallic Roofs Brighter Objects Asphalt Kappa

Kappa because per swimming class pool is so different from the other 1,20 It has had the better result class (color cyan with always rectangular shape). 1,00 1,00 0,80 0,75 0,74 0,71 Kappa 0,60 0,40 0,61 0,59 0,55 0,49 0,47 0,42 0,28 0,20 0,00 Swimming Pool Grass Metallic Roofs Brighter Objects Asphalt Trees Bare Soil Dark Concrete Ceramics Shadow Medium concrete Class

Kappa per class 1,20 Kappa 1,00 0,80 0,60 0,40 1,00 0,75 0,74 0,71 0,61 0,59 0,55 0,49 It has had the worst result. It has been confused with almost all of classes. It will be necessary to redefine the Medium Concrete scope and characteristics. 0,47 0,42 0,28 0,20 0,00 Swimming Pool Grass Metallic Roofs Brighter Objects Asphalt Trees Bare Soil Dark Concrete Ceramics Shadow Medium concrete Class

Kappa per class 1,20 Kappa 1,00 0,80 0,60 0,40 1,00 0,75 0,74 0,71 0,61 0,59 0,55 It has been confused with Dark Concrete, Asphalt and Trees. It was a result of error interpretations in reference polygon. It has been difficult to the interpreter to find visually the color differences among the three classes. 0,49 0,47 0,42 0,28 0,20 0,00 Swimming Pool Grass Metallic Roofs Brighter Objects Asphalt Trees Bare Soil Dark Concrete Ceramics Shadow Medium concrete Class

Kappa per class 1,20 1,00 0,80 1,00 0,75 0,74 0,71 There It was has confused been a with confusion Dark Concrete, between these Asphalt two an classes Trees. It because was a result They of has error the same interpretations color and it in was reference not possible polygon. to use It was the so shape difficult of to the the buildings interpreter (the find segmentation visually the color problem). differences Using among a DSM, the we would resolve three classes. this problem. Kappa 0,60 0,40 0,61 0,59 0,55 0,49 0,47 0,42 0,28 0,20 0,00 Swimming Pool Grass Metallic Roofs Brighter Objects Asphalt Trees Bare Soil Dark Concrete Ceramics Shadow Medium concrete Class

Kappa per class 1,20 1,00 0,80 1,00 0,75 0,74 0,71 This class has been confused with Shadows and specially with Grass. The spectral characteristics were not sufficient to distinguish the classes, because of poor spectral resolution of the sensor. The alternative may be the texture. Kappa 0,60 0,40 0,61 0,59 0,55 0,49 0,47 0,42 0,28 0,20 0,00 Swimming Pool Grass Metallic Roofs Brighter Objects Asphalt Trees Bare Soil Dark Concrete Ceramics Shadow Medium concrete Class

Experiment II N Bairros Selected selecionados Quarters Cidade Jardim Jardim Renata Vila Letônia Vila Acácias Jardim Apolo 500 0 500 Meters it has chosen a specific set of segmentation parameters for each Quarter; I could do re-segmentation operations; The shape of objects are better than the first Experiment ; It has built a specific semantic net for each quarter.

Experimento II Example of a Well - organized Quarter Cidade Jardim

Experiment II N Bairros Selected selecionados Quarters Cidade Jardim Jardim Renata Vila Letônia Vila Acácias Jardim Apolo 500 0 500 Meters Well - organized Quarters The objects are regularly disposed in the urban space; homogeneous size and type of roof material. Jardim Renata, Cidade Jardim e Jardim Apolo Disorganized Quarters heterogeneous size and type of roof material; occurrence of very small objects; the objects are irregularly disposed in the urban space; bigger number of land cover classes than in Well-organized Quarters. Vila Acácias e Vila Letônia

Experimento II Example of a disorganized Quarter Jardim Renata

Exp. II Results (Thematic Accuracy) All of the quarters had better accuracy than the Experiment I expect Vila Letônia Q u ic k b ir d Accuracy Complex B a ir r o s J a r d im R e n a ta C id a d e J a r d im J a r d im A p o lo K a p p a E x a tid ã o G lo b a l 0,7 6 0, 7 9 0,7 4 0, 7 6 0,6 3 0, 6 8 V ila A c á c ia s 0,5 8 0, 6 2 V ila L e tô n ia 0,5 0 0, 5 5

Conclusions The characteristics of cities in development countries brings different challenges in land cover classification; The urban occupation is not well-organized. It is difficult to establish a set of segmentation parameters that works for the whole city. It is recommended to divide de city in homogenous areas (can be the quarters) with specific segmentations and classification parameters. Thus, it will be possible to keep the shape attributes and use re-segmentation operations. Larger test areas demands better software and computers. The poor spectral resolution of the Quickbird sensor could be overcoming by using DSMs to distinguish built classes (Ceramic and Dark Concrete) from classes with the same color (Bare Soil and Asphalt, respectively). Trees X Grass Texture attributes may be a solution.

Thank you very much!

Exp. I Confusion Matrix Reference Classified Polygons Polygons Brighter Objects Ceramics Bare Soil Metallic Roofs Medium concrete Dark Concrete Asphalt Swimming Pool Shadow Trees Grass Total Productor Accuracy User Accuracy Kappa per Class Brighter Objects Ceramics Bare Soil Metallic Roofs Medium concrete Dark Concrete Asphalt Swimming Pool Shadow Trees Grass Total 49 1 0 0 0 0 0 0 0 0 0 50 0 42 6 0 6 1 0 0 6 0 0 61 0 26 30 0 5 1 3 0 0 0 1 66 3 1 0 16 13 4 1 0 6 0 0 44 14 3 2 0 23 1 6 0 3 0 0 52 0 5 3 0 6 16 9 0 11 1 1 52 0 0 0 1 11 3 36 0 3 1 0 55 0 0 0 4 0 0 0 54 2 0 0 60 0 1 0 0 0 4 0 0 52 5 0 62 0 0 2 0 0 0 1 0 17 38 6 64 1 2 7 0 4 0 0 0 9 15 27 65 67 81 50 21 68 30 56 54 109 60 35 631 0,73 0,52 0,60 0,76 0,34 0,53 0,64 1,00 0,48 0,63 0,77 0,98 0,69 0,45 0,36 0,44 0,31 0,65 0,90 0,84 0,59 0,42 0,71 0,47 0,55 0,74 0,28 0,49 0,61 1,00 0,42 0,59 0,75