期刊名称：ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences
印刷版ISSN：2194-9042
电子版ISSN：2194-9050
出版年度：2007
卷号：XXXVI-3/W49A
页码：99-104
出版社：Copernicus Publications
摘要：The process of satellite-based land use and land cover (LULC) mapping often needs to integrate a priori geo-spatial data. A way to consider pre-existing boundaries is per-parcel classification. In the case of crop monitoring on agricultural fields, a per-parcel approach facilitates the classification process by providing pre-defined, ready-to-use boundaries. In other cases, e.g. when using cadastral boundaries, the outlines may not necessarily coincide with LULC information. In this paper we discuss an adaptive per- parcel approach of object generation based on multi-spectral satellite data incorporating the given outlines. The approach differentiates between three cases: (1) a parcel coincides with a homogenous image object; (2) a set of parcels needs to be merged because of homogeneity of the underlying spectral information (3) a single parcel is spectrally heterogeneous, needs to be split and new boundaries are to be generated by object-specific segmentation. The study was carried out in a 3654 km2 sized study area covering the Stuttgart Region in Germany. We used orthorectified, mosaicked SPOT 5 multispectral data (5 m ground sample distance; GSD), co-registered and orthorectified. The digital cadastre data were from 2005. We applied object-based image analysis (OBIA) and used cognition network language (CNL) for modelling objects individually in a semi-automated way. About one fifth of the initial cadastre units have been further subdivided due to internal heterogeneity. But the majority of the units have been merged due to redundancy of the boundaries within. By this, the initial number of units has been reduced to less than one fourth. Expert assessment revealed that more than 96 % of the boundaries dissolved were removed correctly. The accuracy of newly introduced boundaries and also the accuracy of retained boundaries was about 86 %. The result met the demand of the given task, although combining data sets of different scales implied some methodological weaknesses. Overall, high potential of the approach can be attributed to the high degree of automation ensuring cost-efficiency, transferability and compatibility of the results. With regard to related applications, further perspectives of using this approach are given in the conclusion
