A Methodology for the Automatic Generation of Land Use Maps

Transcription

1 456 A Methodology for the Automatic Generation of Land Use Maps Enrico IPPOLITI, Eliseo CLEMENTINI, Stefano NATALI and Gebhard BANKO Abstract Land use maps are usually obtained by a set of semi-automated and manual approaches starting from pre-classified land cover maps. Additionally, other sources of information like census data or cadastral data are also taken into account to obtain the final result. In this paper, we explore a new methodology that automatically produces land use maps by using the topological and semantic information contained in land cover maps. From the observation of existing land use maps of the Land Information System Austria (LISA) project, we infer the rules with which complex objects of land use maps can be built as aggregates of land cover objects. Then, we identify a set of geometric operations that are needed to assess spatial properties and relations among land cover objects. With reference to two use cases, urban settlement and agricultural aggregates, we illustrate the procedures to obtain the land use objects of such cases. 1 Introduction Nowadays, many public and private agencies, like the Environmental Protection Agency of Austria in the Land Information System Austria (LISA) project (WEICHSELBAUM et al. 2009, GRILLMAYER et al. 2010, PRÜLLER et al. 2011), use a set of comprehensive semiautomated and manual approaches, based on classic photo-interpretation techniques, to derive land use information from land cover maps. These methodologies are expensive, time consuming and subjective. In some cases, automatic procedures are applied: for instance, to produce GMES Urban Atlas maps (GMES Urban Atlas 2012), image analysis packages such as ecognition (ecognition 2012) are utilized. Automatic processing techniques may reduce the time employed for manual interpretation, satisfying current demands for continuous and precise data that accurately describes the territory. While land cover is related to the physical characteristics of the earth s surface, land use is related to the socio-economic occupation of the earth s surface: hence, its classification process is more problematic to obtain (DI GREGORIO & JANSEN 2000). Land use is primarily defined in terms of human activities but it can be inferred from the structure of physical components. Spatial patterns and relations (between land cover objects) must be taken in consideration to derive the land use. The field of Object-Based Image Analysis (OBIA) includes contributions towards the object-based classification of satellite images to obtain land cover maps and in fewer cases land use maps (MALINVERNI et al. 2010, NOVACK et al. 2010, THUNIG et al. 2010). Such object-based techniques could be further extended to identify more complex objects starting from the basic knowledge of objects in the classified images. Complex structures (e.g., a nuclear plant or road network) can be identified as the result of the combination of different Jekel, T., Car, A., Strobl, J. & Griesebner, G. (Eds.) (2012): GI_Forum 2012: Geovizualisation, Society and Learning. Herbert Wichmann Verlag, VDE VERLAG GMBH, Berlin/Offenbach. ISBN

2 A Methodology for the Automatic Generation of Land Use Maps 457 objects or of groups of objects (BALTSAVIAS 2004). This requires the study of the spatial and semantic relations among the objects and their definition in computer language (LIU et al. 2008). Early work in this area is described, e.g., in (BARNSLEY et al. 2001), defining a Structural Analysis and Mapping System (SAMS) that is able to understand urban land use in terms of structural composition of the land cover objects. More recently, the work in (HUSSAIN et al. 2007) describes a pattern recognition methodology to automatically classify urban structure by using their morphological properties and topological spatial relations. While the latter methods explicitly search for spatial relations and try to understand the spatial structure among land cover objects, the proposal of (WIJNANT & STEENBERGHEN 2004) represents a more black-box oriented method that statistically examines the land cover information to determine the land use of the parcel. The overall aim of the research described in this paper is to formalize a methodology that automates the land use classification for the LISA project. The assumption underlying our approach is that land use objects can be identified on the basis of semantic and spatial information about land cover objects. A distinguishing aspect of the proposed methodology is that land cover maps and the output land use maps are stored in a vector data model. We suppose to use OGC Geometry Object Model (OGC 2011) to describe the objects geometry. Various computational geometry algorithms are used for checking geometric properties and relations between land cover objects and for transforming geometries. The remainder of the paper is organized as follows: two LISA use cases on residential urban areas and agricultural complex objects are described in section 2. The proposed methodology that automates land use classification is outlined in section 3. In section 4 our methodology is applied to the use cases previously selected. Section 5 presents a possible integration of methodology in a web-based GIS architecture. The paper ends with some considerations and proposes possible further developments. 2 LISA Use Cases The objective of LISA is to achieve a consensus on a new Austrian land information system and demonstrate its benefits offering improved spatial and thematic content. LISA is designed to serve common land monitoring needs providing information on the status quo and the changes occurring in Austria s landscape. We identified two use cases, where the application of our methodology will reproduce and, in some cases, improve upon the current techniques to build land use objects in LISA by exploiting the knowledge about land cover: these use cases are related to residential urban settlements and agricultural land use categories. 2.1 Urban Settlement use case The Urban Settlement use case aims at defining urban land use objects such as residential areas (Figure 1a). The candidate land use object is derived using roads as basic delimiting factor: only the important roads are used for this. Smaller roads, which are only used as access roads to buildings, are not used as boundary of land use object s geometry but are considered inside the land use object. Buildings are judged based on their size (only small/medium buildings can represent residential usage). The urban settlement includes also other neighbouring land cover objects, such as green areas and groups of trees (usage:

3 458 E. Ippoliti, E. Clementini, S. Natali and G. Banko garden). A further aspect to distinguish residential areas compared to industrial areas could be the size of the buildings and the proximity to constructed areas non belonging to a road network (usage: parking lots). 2.2 Agriculture use case Like the previous use case, the candidate agricultural objects are, in a first step, differentiated using roads, forest borders and rivers as basic delimiting factor. A further classification criterion is the orientation and shape properties of objects: elongated objects of about the same orientation are merged together to form a single agriculture complex object (Figure 1b). (a) Fig. 1: (b) Land use objects: (a) candidate residential areas; (b) candidate agricultural areas 3 Methodology In our methodology we suppose to consider as input a pre-classified map containing only simple features belonging to land cover classes. The expected result of the automatic procedure to generate land use from land cover is shown in Figure 2b. (a) Fig. 2: (b) Generation of land use map from land cover map: (a) land cover map contains simple objects such as buildings, tree areas, herbaceous areas and other constructed areas (all represented in the picture); (b) land use map contains complex objects such as residential areas (in the picture), industrial areas and agricultural areas

4 A Methodology for the Automatic Generation of Land Use Maps 459 Land cover maps used for testing the proposed methodology are based on LISA data model. LISA defines a set of land cover classes, like building, other constructed area, tree, surface water and a set of land use classes like urban settlement, agricultural block, forest block to classify geographic features. In Figure 2(a), a land cover map is shown: generally a land use object can be represented as an aggregation of one or more land cover objects. The proposed methodology consists of four phases: (1) land use object definition; (2) geometric operators identification; (3) functions definition; (4) validation. The land use object definition phase aims at defining geometric properties and semantic relations that characterize land use objects. Specifically, for each land use object we need to define a set of land cover objects with their geometric attributes and the semantic and geometrically relations among them: for example, the reference distance between land cover objects, the reference size for land cover objects, their reference shapes, and the occurring topological relations. Also, for each land use object, we need to define the required minimum mapping unit (MMU) such that the object can be considered valid. The geometric operators identification phase aims at identifying the geometric operators (via computational geometry algorithms) that allow us to compute and/or to verify geometric properties and spatial relations identified in the previous phase. The set of geometric operators can be subdivided into two subsets: operators that evaluate the geometric properties and operators that transform the geometries. Examples of geometric operators belonging to the first subset are shape and size operators, distance operators, and topological operators. Geometric operators belonging to the second subset are cartographic generalization operators (such as merging, simplification, aggregation, fusion), container operators (e.g., convex hull, minimum bounding rectangle, minimum bounding circle), and morphological operators (e.g., buffers). The functions definition phase is needed to define a set of parameterized functions that, performed in a given sequence, permit to identify land use objects. Typically, the first functions to be applied are those that aggregate and group land cover objects into intermediate categories; subsequent functions refine land use objects with cartographic generalization operators and/or morphological operators. Finally, the validation phase aims at testing functions formalized in the previous phase on real datasets. Measures of accuracy must be performed comparing the obtained results with existing land use maps. 4 Application to Use Cases Residential areas can be recognized by analysing the shape and size of buildings and the spatial relations among them. In order to implement Urban Settlement use case, the sequence of functions that must be carried out is the following: (1) Object Aggregation; (2) Object Grouping; (3) Object Refinement; (4) Object Size Check. The Object Aggregation function aims at creating the urban settlement aggregates putting together the land cover objects that semantically belong to specific land cover classes ( building, tree, bushes and herbaceous ), that are dimensionally comparable and not too big or small with respect to a reference size, and that satisfy the topological relation touch. The touch relation is satisfied when two objects geometries have at least one boundary point in common but no interior points. The function starts considering a single land cover object that belongs to

5 460 E. Ippoliti, E. Clementini, S. Natali and G. Banko building class and that is small/medium in size with respect to a given reference size. It continues by aggregating the given building object with other land cover objects that belong to specific classes like tree, bushes and herbaceous (usage: garden), that are comparable in size and that are in touch (Figure 3a). (a) (b) (c) Fig. 3: (d) Urban Settlement aggregates construction: (a) the process starts from building (in dark) situated in the upper left corner; b) iterative aggregation; (c) first urban settlement aggregate (id=1) constructed (object geometry is a Polygon); (d) final result (five aggregates constructed: ids=1,2,3,4,5) The process of aggregation continues by adding other land cover objects that satisfy previous constraints with the constructed object so far until the urban settlement aggregate is completely surrounded by land cover category other constructed area (Figures 3(b) and 3(c)). The previous steps are repeated starting from another building object that is not included in any urban settlement aggregate already constructed (Figure 3(d)). The Object Aggregation function uses the size operator to compute the land cover objects size, the touches operator to check the topological relation touch, and the merge operator to aggregate land cover objects geometries that are in touch. Fig. 4: The Object Grouping function constructs one cluster (object geometry is a MultiPolygon)

6 A Methodology for the Automatic Generation of Land Use Maps 461 The Object Grouping function groups together the identified urban settlement aggregates. Clustering algorithms, based on boundary or centroid distance, are used to identify urban settlement aggregates to group in a single cluster (Figure 4). This function uses the distance operator to compute the distance among objects and the union operator to group objects geometries into a single geometry collection. The Object Refinement function is applied on each cluster previously constructed. It involves two steps: the first step consists in fusing the object geometric components that are closer than the identified distance, by making the convex hull of neighbouring parts of boundaries (Figure 5a), while the second step performs a regularization of the object shape, by filtering small irregularities (filling small concavities) and by replacing irregular sides with straight edges (Figure 5b). Finally, the Object Check Size function carries out a validation of the land use objects constructed so far. Only land use objects that have an area larger than the MMU of the output data set are considered valid. The MMU established for urban settlement objects is 1000 m 2. (a) (b) Fig. 5: Residential areas construction: (a) fusion transforms a MultiPolygon (A1, A2, A3, A4 and A5, see Fig. 4) in a Polygon (A); (b) shape regularization Regarding the Agriculture use case, a very similar approach is performed in our methodology. The sequence of functions that must be carried out is the following: (1) Object Aggregation; (2) Non-Oriented Object Inclusion; (3) Object Check Size. The Object Aggregation function finds out agricultural aggregates, joining land cover objects that semantically belong to specific crop classes, that satisfy the topological relation touch, and that have similar (and well defined) orientation (Figure 6a). In this case the Object Aggregation function uses, in addition to touches and merge operators, new geometric algorithms to compute the shape orientation of land cover objects. The Non-Oriented Object Inclusion function serves to aggregate the land cover objects with agricultural field label but with non-clear orientation (e.g., object of square or circle shape) with the agricultural aggregates identified using the previous function with which they have the highest number of touch relations and with which they have the same crop class (Figure 6b). The last consideration is useful to preserve a compact form of agricultural aggregates. Finally, the Object Check Size function carries out a validation of the land use objects constructed so far. Only agricultural aggregates that have an area greater than the MMU of the output data set are considered valid. In this case the value of MMU for valid agricultural land use objects is established at 5000 m 2.

7 462 E. Ippoliti, E. Clementini, S. Natali and G. Banko (a) Fig. 6: (b) Agricultural areas construction (partial): (a) agricultural aggregates identified (six aggregates constructed: ids=1,2,3,4,5,6); (b) refinement applying the Non- Oriented Object Inclusion function (small non-oriented object is aggregate to the first agricultural complex object, id=1) 5 Software Prototype This section describes a software prototype being implemented to permit to perform a second level classification (from land cover to land use) through the application of functions among objects belonging to the first level of classification, provided as input to the system. We suppose to implement this software system in a web-based architecture. Fig. 7: Software prototype architecture Figure 7 describes the system architecture in terms of its main functional components. The software prototype consists of three main logic layers: the database layer, where the land cover maps provided by the user are stored and accessed by the server side for processing, where the results of the processing (land use maps) are stored, and where the definitions of land use objects (as sequence of functions with their respective configuration parameters, named Complex Objects Definitions, CODs) are stored; moreover the database layer hosts the so-called tiles database, a file-system data structure created by a web map tile service to speed up data visualization;

8 A Methodology for the Automatic Generation of Land Use Maps 463 the server / processing layer that contains all the processing modules for both data ingestion / export and processing; the GUI layer that is the user interface through which each user provides its processing inputs, visualize the results and export them into a map (vector layer). Following sections summarize the characteristics of each layer. 5.1 Database layer The main functional scope of the database layer is the storage of the dataset provided in input by the user and storage of temporary and final land use objects identified applying, in a given sequence, specific geometric-semantic functions (the user-defined rules, stored as CODs). Moreover the database layer shall host the stored land use objects definitions (CODs) for a successive retrieval / application / modification. Specific I/O modules to permit the mapping of the input datasets to the internal data model and of the data model toward the output data formats must be developed. Finally, in order to speed up data visualization, a tiles database is hosted and updated each time a new complex object is created. The database layer has an interface toward the server / processing layer (for I/O and R/W purposes during the land use object definition) and toward the GUI layer for original data and results visualization. 5.2 Server / processing layer Fig. 8: Software modules in the three layers The server / processing layer is the core of the system and hosts all the processing modules for both I/O purposes and land use objects definition purposes (Figure 8). The layer has to perform the three main functionalities: to manage I/O operations, permitting loading of a new image on the database and permitting exporting land use layer; to manage land use object definitions (retrieval / storage); to create specific queries / operations on the database to identify land use objects on the input land cover map. This server / processing layer has an interface toward the database layer as well as toward the GUI layer. 5.3 GUI layer The GUI layer is the direct interface toward end user. The GUI shall allow the user to: load a land cover dataset into the system for processing; visualize the input dataset;

9 464 E. Ippoliti, E. Clementini, S. Natali and G. Banko load an existing land use object definition to be applied to the ingested dataset; define and store specific rules combining geometric operators and functions for land use object identification; visualize the result (identified land use objects); export the results of the elaboration as new set of files. The GUI layer has direct interface toward the database layer for the visualization of the original dataset and of the identified land use objects, making use of OpenLayers technologies; moreover the GUI layer has interfaces toward the server side for the data I/O and for the definition of the land use objects identification rules (land use objects properties). 6 Conclusions and Further Work In this study, we outlined a new methodology to automate the identification of land use objects from land cover maps in the LISA project. As use cases, we applied it to the identification of residential areas and agricultural areas. The distinguishing aspect of our approach is that we use only land cover information without consider ancillary data. Also our methodology is applicable on land cover maps stored in a vector data model. This implies the use of OGC Geometry Object Model to describe the object s geometry and the use of various computational geometry algorithms to check geometric properties, to compute spatial relations between land cover objects, and to transform geometries. As a further development of this work, we will apply the proposed methodology to other LISA use cases, such as the identification of industrial and commercial areas and the identification of river systems, which are complex objects composed of a river and other small land cover objects close to it. Also, we plan to implement the proposed software prototype to validate our approach with LISA end-users. Acknowledgements The research described in this paper was supported by European Space Agency (ESA), trough the Support To Topology (STO) project ( References BALTSAVIAS, E. P. (2004), Object Extraction and Revision by Image Analysis Using Existing Geodata and Knowledge: Current Status and Steps Towards Operational Systems. ISPRS Journal of Photogrammetry & Remote Sensing, 58, BARNSLEY, M. J., MØLLER-JENSEN, L. & BARR, S. L. (2001), Inferring Urban Land Use by Spatial and Structural Pattern Recognition. In: Remote Sensing and Urban Analysis (Ed. by J.-P. Donnay, M. J. Barnsley & P. A.Longley). Taylor & Francis,

10 A Methodology for the Automatic Generation of Land Use Maps 465 DI GREGORIO, A. & JANSEN, L. J. M. (2000), Land Cover Classification System (Lccs): Classification Concepts and User Manual, FAO. X0596E/X0596e00.HTM. Rome. ECOGNITION (2012), GMES URBAN ATLAS (2012), GRILLMAYER, R., BANKO, G., SCHOLZ, J., PERGER, C., STEINNOCHER, K., WALLI, A. & WEICHSELBAUM, J. (2010), Land Information System Austria (Lisa) Objektorientiertes Datenmodell Zur Abbildung Der Landbeckung Und Landnutzung. In: Angewandte Geoinformatik 2010 Beiträge Zum 22. Agit-Symposium (Ed. by J. Strobl, T. Blaschke & G. Griesebner). Wichmann, HUSSAIN, M., DAVIES, C. & BARR, R. (2007), Classifying Buildings Automatically: A Methodology. GISRUK 2007: Proceedings of the Geographical Information Science Research UK 15th Annual Conference. A. C. Winstanley. Maynooth, Ireland, 11th-13th April 2007, LIU, Y., GUO, Q. & KELLY, M. (2008), A Framework of Region-Based Spatial Relations for Non-Overlapping Features and Its Application in Object Based Image Analysis. ISPRS Journal of Photogrammetry & Remote Sensing, 63, MALINVERNI, E. S., TASSETTI, A. N. & BERNARDINI, A. (2010), Automatic Land Use/Land Cover Classification System with Rules Based Both on Objects Attributes and Landscape Indicators. GEOgraphic Object-Based Image Analysis GEOBIA Ghent, Belgium, geobia.ugent.be/proceedings/html/papers.html. NOVACK, T., KUX, H. J. H., FEITOSA, R. Q. & COSTA, G. A. (2010), Per Block Urban Land Use Interpretation Using Optical Vhr Data and the Knowledge-Based System Interimage. GEOgraphic Object-Based Image Analysis GEOBIA Ghent, Belgium, geobia.ugent.be/proceedings/html/papers.html. OGC (2011), Geometry Object Model. OpenGIS Implementation Specification for Geographic information Simple feature access Part 1: Common architecture, PRÜLLER, R., GRILLMAYER, R., BANKO, G., MANSBERGER, R., STEINNOCHER, K., STEM- BERGER, W., WALLI, A. & WEICHSELBAUM, J. (2011), Nutzen Von Innovativen Technologien Für Eine Flächendeckende, Flexible Landbeobachtung Österreichs. In: Angewandte Geoinformatik 2011 Beiträge Zum 23. Agit-Symposium (Ed. by J. Strobl, T. Blaschke & G. Griesebner). Wichmann, THUNIG, H., WOLF, N., NAUMANN, S., SIEGMUND, A. & JÜRGENS, C. (2010), Automated Lulc Classification of Vhr Optical Satellite Data in the Context of Urban Planning. GEOgraphic Object-Based Image Analysis GEOBIA Ghent, Belgium. geobia.ugent.be/proceedings/html/papers.html. WEICHSELBAUM, J., BANKO, G., HOFFMANN, C., RIEDL, M., SCHARDT, M., STEINNOCHER, K., WAGNER, W. & WALLI, A. (2009) Land Information System Austria (Lisa): Bedarfsgerechte Landnutzungsinformationen für die Öffentliche Verwaltung. In: Angewandte Geoinformatik 2009: Beiträge Zum 21. Agit-Symposium (Ed. by J. Strobl, T. Blaschke & G. Griesebner). Wichmann, WIJNANT, J. & STEENBERGHEN, T. (2004) Per-Parcel Classification of Ikonos Imagery. 7 th AGILE conference on Geographic Information Science. Heraklion, Greece,