A Bilateral Filter Based Post-Processing Approach for Supervised Spectral-Spatial Hyperspectral Image Classification

To effectively improve the performance of representation based classifier, a spatial spectral joint classification post-processing approach is proposed, based on the application of edge preserving BF (Bilateral Filtering) method. The proposed framework includes two key processes: (1) the classifier (such as SRC, CRC, or KSRC) based on sparse representation of each pixel is used to obtain soft-classified probabilities belonging to each information class for each pixel; (2) spatial spectral joint BF for the soft-classified probabilities map. It is aimed to integrate context-aware information for each pixel class labels. Under the spatial guidance image, extracted from the three principle component, a BF is employed to get the refined probability maps. The BF considers not only the spatial distance but it also considers the image context-aware distance which significantly improves the classification results. Finally, the class label is obtained by choosing the maximum probability criteria. The experimental results on three benchmark hyperspectral data sets showed that the ＂local smoothing＂ is efficient and has a potential to achieve high classification accuracy. All the algorithms are implemented with equal number of labeled samples and comparative results are presented in terms of visual classification map and numerical classification results. The major advantages of proposed method are: it is simple, noniterative and easy to implement. Hence, the advantages lead to significant usage in real applications.