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：2000
卷号：XXXIII Part B7(/1-4)
页码：1423-1430
出版社：Copernicus Publications
摘要：Large amounts of remotely sensed data are collected from airborne platforms with spatial resolutions from 2 m down to 15 cm and similar satellite systems are coming on line. Conventionally they have been analysed by manual airphoto interpretation techniques, which are time-consuming, subjective and expensive, therefore it is likely that semi- automated approaches will be attempted. Data with a fine spatial resolution would appear to be a major advantage for mapping as the proportion of mixed land cover pixels is reduced. It does, however, present a new set of problems related to how these types of data should be analysed using automated and semi-automated techniques. At finer spatial resolutions the data recorded by an instrument begin to cross a scale boundary where it is related not to the character of objects or areas as a whole, but to components of them. For instance, with 30 m Landsat TM data, pixels within a suburban area will each represent a mix of buildings, vegetation and bare ground etc. that would generate a reasonably unique spectral signature and they could usually be classified by conventional semi-automated techniques. As the spatial resolution gets finer pixels begin to represent components within the suburban area, such as buildings, gardens, paved areas or mixtures of these. This increases the number of classes that must be mapped, reduces their spectral separability and increases the complexity of the result. This paper addresses two problems associated with the automated analysis of fine spatial resolution data. Firstly, how to classify the images into meaningful surface features when training areas would be difficult if not impossible to find. Secondly, how to integrate the detailed information returned by the classifier to more usable classes and more appropriate scales. To classify fine spatial resolution data it will be necessary to identify a set of scene components which manifest themselves at that particular spatial resolution. The scene components must then be labelled as real world features after an assessment against the spatial resolution. Recent advances in integrated Geographical Information Systems have allowed the development of automated procedures for analysing remotely sensed data on a per-parcel basis, rather than the conventional per-pixel basis. This allows aggregation of the scene component information within a region, the consideration of context, both within the region and beyond it, and the application of knowledge-based rules at the parcel level to classify the parcels into meaningful land cover classes. A case study is provided which highlights the problems associated with classifying fine spatial resolution images and also describes the parcel-based approach to the classification with scene components. The case study used data from the High Resolution Stereo Camera-Airborne (HRSC-A) system with an original spatial resolution of 15 cm to map land cover type within forest stand parcels in the Tharandter Forest, Germany. Although this methodology for image classification makes no radical advances in terms of correspondence accuracy compared to conventional per-pixel classification of coarse spatial resolution data it does hold its own and provides a much richer set of results and an opportunity for further analysis
关键词：Fine spatial resolution; classification; per-parcel; scene components; HRSC-A
