An Object-oriented Classification Approach for Analyzing and Characterizing Urban Landscape at the Parcel Level Weiqi Zhou, Austin Troy& Morgan Grove University of Vermont 2006 BES Annual Meeting
Outline Background Research Objectives Methodology Results Conclusions
Background Questions & Motivations – What are the ecological and social consequences associated with urban expansion? The structure of the urban landscape must be first quantified and understood. – Land cover data at the parcel level is highly desirable because land management decisions are commonly made at the individual household level.
Background Data – High-resolution imagery is needed for mapping detailed urban land cover; Coarse spatial resolution imagery (e.g., MODIS, Landsat TM) is insufficient; Increased availability of high-resolution satellite imagery, such as IKONOS and QuickBird, as well as digital aerial imagery
900m 2
0.61m panchromatic QuickBird imagery 1m IKONOS data 0.6 m Emerge data
Background Classification approaches – Visual interpretation is time consuming and expensive; – Conventional pixel-based methods are inadequate; – Object-oriented classification Segment images into objects; Incorporate shape, spatial relations, and reflectance statistics, as well as spectral response; Similar to visual interpretation, but has the advantage of minimal human interaction.
Objectives Develop an object-oriented approach for classifying and analyzing the complex mix of vegetation and development in urban landscapes at the parcel level using high-resolution digital aerial imagery and LIDAR data. Apply this approach to map and inventory the land cover in the Gwynns Falls watershed.
Study sites
Methods Data collection and preprocessing – Emerge aerial imagery Color infrared image, collected in 1999, pixel size of.6m – LIght Detection And Ranging (LIDAR) data Collected first and last returns in 2002; Interpolated into raster data with pixel size of 1m; – Points from first returns Surface cover DSM – Points from last returnsGround DSM Surface cover height = surface cover DSM – Ground DSM;
Methods Data collection and preprocessing – Parcel boundary data Helped segment objects; Used as functionally meaningful geographic unit by which to summarize landscape composition. – Building footprint data
Methods Image Segmentation – Fractal net evolution approach, embedded in eCognition; – Segmented the image into 3 levels of objects; Level 1: objects were considered to be internally homogeneous; Level 2: classification-based fusion Level 3: Parcel
Level 1 segmentation Level 1
The roof was segmented to 3 object primitives
Level 2 segments Level 2
Level 1 Details Level 1
Level 2 details Level 2
level3 Level 3
Methods Fuzzy classification – Land cover types of interest Building Coarse textured vegetation (trees & Shrubs) Fine textured vegetation (herbs & grass) Pavement Bare soil
Methods Fuzzy classification – Class hierarchy A class hierarchy was developed, and a number of rules were created to classify each object into one of the 5 classes at the most disaggregated spatial level.
Methods_Hierarchy Buildings NonBuildings Building footprint data Brightnes s Context& Texture LIDAR data Context&NDVI Shaded Fine Vege Shaded Pavement Context Shaded Coarse Vege Shaded Buildings Shaded short Shaded tall LIDAR data Fine Vegetation Coarse Vegetation NDVI Missing Building Bare Soil Pavement NonVegetation Vegetation NonShaded Shaded
Results_GF
Results
Accuracy Assessment Overall Accuracy: 92.3% (n=350)
Results_Parcel_Impervious
Result_parcel_FineVege
Result_parcel_LawnGreenness
Result_parcel_TreeCanopy
Conclusions (1) The OO classification approach proved to be effective for classifying urban land cover from high-resolution multispectral imagery; Ancillary data, such as LIDAR can greatly improve the classification;
Conclusion (2) The OO approach using parcels as pre- defined patches provides a convenient and effective way for integrated research to incorporate biophysical and social factors, especially the research on relationships between household characteristics and structures of urban landscapes.
Acknowledgements This research was funded by the Northern Research Station, USDA Forest Service and the National Science Foundation LTER program (grant DEB – ). Thanks to a lot of BES people.
Questions?