Efficient marker controlled watershed algorithm for classification of hyperspectral images

Hyperspectral image classification becomes a prominent topic in remote sensing. Hyperspectral image provides in detail spectral and spatial information about earth surface object. With the help of spectral and spatial information, it is highly possible to distinguish spectrally similar objects. But hyperspectral images are with hundreds of spectral bands which lead to lacking the availability labeled samples and high cost of computation. To identify earth surface objects accurately, both issues such as large number of spectral channels and limited availability of training samples should be addressed properly in classification tasks. In this paper, we divide a large dataset into regions with watershed segmentation algorithm and then conducting coarse to fine hypergraph construction. In the first layer, first we compute the pairwise relevance, which fed to the second layer from which hypergraph is constructed in the second layer. Semisupervised learning is employed on hypergraph to obtain a final classification map. In our proposed system segmentation helps to reduce the computation burden while coarse to fine hypergraph based learning helps to tackle issues such as high dimensionality and few training samples.