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：2008
卷号：XXXVII Part B7
页码：641-650
出版社：Copernicus Publications
摘要：The information in a pixel of satellite data within the instantaneous Field of View (IFOV) of the sensor is a mixture of different land cover types, and the individual land cover components can be estimated using soft classification techniques. However these techniques do not account for the spatial distribution of the class proportions, the information itself has a great relevance. In this study Markov Random Field (MRF) based Super Resolution Mapping (SRM) with certain modifications have been analysed for its performance with respect to the linear unmixing technique applied on hyperspectral data. This testing was carried out over a heterogeneous urban environment which was defined by the Vegetation, Impervious surface and Soil (V-I-S) model that has been used as an accepted alternative in characterising the urban land cover components. Linear unmixing technique with a hyperspectral remote sensing image (Hyperion) has been used to generate fractions according to the spectral variability of the V-I-S classes. Modified MRF based SRM technique was applied on IKONOS, ASTER MSS and Landsat ETM+ images with markedly different spatial and spectral resolutions. Reference map for the validation were created from the IKONOS MSS image using hard Maximum likelihood classification. The super resolution maps which contain the spatial information were again turned in to fractions representing each class (V-I-S). Next the results of MRF based SRM technique and the linear unmixing technique were validated using three measures of accuracy with respect to the reference fractions of the IKONOS image, the Area error proportion (AEP), Root Mean Square Error (RMSE) and the Correlation Coefficient (CC). The accuracy statistics for the Optimized Super Resolution Map (OSRM) fractions and the reference fractions showed a higher CC value in the range of 0.7 with respect to the linear unmixing fractions which lies in the range of 0.5. Results were justified by the overall RMSE and AEP values dropping form 0.7 to 0.5 and 1.8 to 1.4 respectively. In the case of ASTER visible to infrared (VNIR) image the correlation of the OSRM with the reference again showed a higher value in the range of 0.7 than the linear unmixing results which showed a correlation in the range of 0.5. Here it has been seen the overall RMSE and the AEP values were dropped for the SRM than the linear unmixing results from 0.7 to 0.6 and 1.8 to 0.6 respectively. In addition the results for the ASTER short wave infrared (SWIR) image and the Landsat image also followed a same trend which finally envisaged the improved subpixel representation of these land cover classes with the use of MRF based SRM techniques than the linear unmixing technique. The contextual refinement brought in by the MRF based SRM technique can produce accurate land cover components at a sub-pixel level even in a heterogeneous urban environment.
关键词：Hyperspectral Images; Spectral Mixture Analysis; Markov Random Field; Simulated annealing; Super Resolution ; Mapping
