Supervised Sparse-Representation Classification on Hyperspectral Images Using the City-Block Distance to Improve Performance

Hyperspectral Images’ data structure contains information from spectral ranges far beyond the limits of conventional, visible light imaging devices. This additional information comes at the cost of having unwieldy large image sizes, which presents an important computational resource consideration for many applications. Recent studies have shown that the usual large amount of redundancy in Hyperspectral Images can be discarded to avoid classification noise by means of sparse signal processing principles. To improve on classification performance, class-dependent Sparse-Representation classification (cd-SRC) algorithm includes Euclidean distance information between the sparse representation of a sample pixel to be classified and the training classes. The current work describes the use of Manhattan (or City Block) distance for improving the cd-SRC classification procedure. Our results show that classification performance using Euclidean Distance is comparable to the Manhattan Distance metric, which requires a smaller number of significantly less computationally expensive operations. In addition, we show that sparse representation classification has a significant advantage in classification performance compared to established Hyperspectral Image classification algorithms, such as the well-known Minimum Euclidean Distance and Support Vector Machine classifiers.