Per-pixel and object-oriented classification methods for mapping urban features using Ikonos satellite data

摘要

Urban planning requires timely acquisition and analysis of spatial and temporal information for making informed decisions. Whilst spectral analysis of images has yielded satisfactory results, they may not be enough to extract urban features from very high resolution (VHR) satellite data such as Ikonos. A combined spectral and spatial approach may be useful to map urban features particularly those with low spectral separability. The paper describes an approach using both per-pixel and object-based classification methods for mapping urban features from VHR satellite data. We tested the suitability of Ikonos satellite data for mapping urban features at a planning scale in near-real time. Parametric per-pixel supervised (maximum likelihood) classification methods are used in combination with object-based classification methods to map urban features over New York City. We employed a combination of spectral, spatial attributes and membership functions for mapping urban features. Accuracy assessment was carried out using ground truth data acquired from field surveys and from other reliable secondary data sources. Whilst the per-pixel approach produced reasonable overall accuracy, specific classes such as white roof and vegetation registered low user's accuracy (79.82 and 70.07) respectively. We were able to improve the accuracy of these two classes by using an object-oriented classification method further to 89% and 97%. The combined approach using per-pixel and object-oriented classification methods may prove useful in the analysis of VHR satellite data like Ikonos, Quickbird, since it results in higher per class accuracy. In this study different urban classes were extracted that can be exported into GIS for further analysis and modeling. Mapping output generated in this study may be beneficial to planning, environmental and emergency services that depend on current geospatial information either for mapping land use changes, or for rapid updating of current maps and spatial information, and management of resources in near real-time. Given the high spatial accuracy, but limited spectral resolution of Ikonos data, we recommend a combined classification approach for extracting sub-pixel urban features.