Dimensionality Reduction of Hyperspectral Images Using Pixel Block Distance for Efficient Classification

Hyperspectral imaging provides useful information in the field of satellite imaging. The extensive spectral and spatial data contained within hyperspectral images (HSI) needs to be analyzed for retrieving the insights about the geographical details. The redundancy in spectral bands results in high dimensionality. The high dimensionality of HSI results in increased computational complexity. This subsequently influences the classification accuracy. Consequently, dimensionality reduction (DR) of HSI is essential before the classification. State-of-the-art DR techniques fail to recognize the nonlinearity in hyperspectral data. The proposed distance measure utilizes information derived from both spatial and spectral domains. The existing DR methods can be included with the proposed distance measure to address the issue of nonlinearity in HSI data. The tests are carried out on DR methods included with the proposed distance measure to assess the results of classification utilizing Support Vector Machine (SVM). Results of classification show that DR techniques incorporated with proposed measures address the nonlinearity present in HSI data.