In order to enhance the accuracy of hyperspectral remote sensing classification, a classification method based on SVM with end-member extraction is presented. Firstly, the end-members are extracted using pure pixel index approach, and then the ground target is identified based on the spectral feature fitting , followed by the spectral classification of the hyperspectral remote sensing images with the Support Vector Machines. The experiment results indicated that the validity and efficiency of our method are more accurately than the traditional SVM solutions which simply use the regions of interest selected from image as the training samples.
The anomaly of surface object associated with underground nuclear explosion (UNE) was regarded as detected target using remote sensing image, because of its unknown geometrical and spectral feature, it is difficult to detect anomaly using general methods of change detection based on gray and texture feature using remote sensing data. A multi-feature joint matching method of anomaly detection was developed based on spectral matching. Firstly, taking pixel spectral curve of image acquired before the UNE as reference spectral vector, and the one acquired after the UNE as matched spectral vector, then the similarities of spectral shape, feature and value were calculated through spectral matching algorithm. Secondly, each similarity above was normalized and weighted by appropriate coefficient; the multi-feature joint matching result was produced by accumulating the weighted similarity. Finally, appropriate segment method was chosen to distinguish target from background, and the pixels with high unmatched degree (namely anomaly) were detected. We took some UNE events for example, the anomaly of the UNE was detected effectively, and the problem of undetected anomaly just using single spectral matching algorithm is well resolved, the result also suggests the method be in good generality.