Object-based approaches to image classification for hyperspatial and hyperspectral data.

摘要

The prime objective of this research is to develop a suitable object based classifier for detailed land use/land cover classification (LULC) of remote sensing data with high spatial and spectral resolution. Owing to technical limitations, remote sensing data were available either at high spatial resolution (4 bands) but not with combination of both until recently. Processing of the high spectral resolution imagery for LULC classification was predominantly pixel based due to the lack of sufficient spatial resolution for identifying individual objects. For high spatial resolution imagery, object based analysis was devised that performed classification at individual object level. But detailed object classification was restricted due to the limitations in the spectral resolution. Recently, the advancements in remote sensing technology have made hyperspectral imagery with high spatial resolution available that permits object-based processing of these datasets for a detailed LULC classification. However, currently available object-based classifiers are only modifications of the pixel based classifiers developed for multispectral data. They are either parametric in nature with the assumption of Gaussian distribution and/or do not completely exploit the rich spectral information available in the hyperspectral imagery. This research proposes a supervised non-parametric fuzzy classifier that performs classification based on the object-level distribution of reflectance values. A fuzzy Kolmogorov-Smirnov based classifier is proposed that performs an object-to-object matching of the empirical distribution of the reflectance values of each object and derives a fuzzy membership grade for each class without any distributional assumptions. This object based classification procedure was tested for its robustness on three different sensors with varying combinations of spectral and spatial resolutions. General land use/land cover classifications as well as detailed urban forest tree species classifications were performed to test the performance of the classifier. The results for the two study areas show that the proposed classifier consistently achieves high accuracies, irrespective of the sensor, and also demonstrates superior performance in comparison to other popular object and pixel-based classifiers.