Comparison of Two Algorithms for Land Cover Mapping Based on Hyperspectral Imagery

This paper presents an analysis of hyperspectral image data, carried out using two approaches followed by a comparison of the two. The hyperspectral image dataset used in this analysis corresponds to a strip along the North Eastern region of Sri Lanka, obtained by the Earth Observing (EO-1) satellite’s Hyperion sensor. Mapping land-cover using hyperspectral imagery makes it possible to obtain finer details of land-cover, which are not obtainable using RGB images. Therefore, hyperspectral imagery could be used to obtain useful information for natural resource location and ecosystem service management, assessing the human induced and natural drivers of changes in land, foliage or water bodies and even in the identification of fine details such as the distribution of minerals in an area before doing a ground survey.
The two algorithms discussed in this paper, initially represent each pixel as a point in a high dimensional space of which the dimensions represent each band of wavelength and subsequently follows two unique approaches to cluster the points (pixels) in a reduced dimensional space. The first algorithm discussed in this paper employs Principal Component Analysis (PCA), Fisher Discriminant Analysis (FDA) and Spectral Clustering in a logical sequence, while the second uses PCA along with concepts of Euclidean geometry. The pixels belonging to each cluster were labeled under ‘soil’, ‘foliage’ or ‘water bodies’, with the aid of the k-means algorithm and the hyperspectral image data of the training set obtained with the aid of Google Maps. The classification process is followed by a comparison of the two approaches employed. Conclusively, the two approaches discussed have their own pros and cons, whilst providing promising results. Hence, both algorithms could be used appropriately based on the application.