Supervised classification of snow cover using hyperspectral imagery

Snow cover classification maps are a significant input in snowmelt runoff models for understanding the hydrological processes. While hyperspectral remote sensing provides significant opportunities in the assessment of land cover features, it is yet underexplored in the snow-covered areas. Dimensionality reduction is extremely significant due to a large number of spectral bands in hyperspectral imagery and also because of relatively small number of training pixels in difficult snow covered terrains. Hyperspectral band selection is indeed a key preliminary step for classification. In this study, a hyperspectral band selection technique is proposed which utilizes the mutual information (MI) between different spectral bands and the reference data. Different variants of the proposed method were experimented, which includes pre-clustering of bands before the computation of MI. The paper emphasizes computationally efficient techniques for the selection of optimal bands in the supervised classification of hyperspectral imagery corresponding to snow-covered mountainous regions. The proposed methods are evaluated with a data set corresponding to the Solang, Himachal Pradesh, India. The methods from the proposed approach are evaluated against state of the art techniques based on statistical accuracy and computational time.