First steps in the dimensionality reduction of hyperspectral images of real-world scenarios based on curve representation

Dimensionality reduction (DR) has been used in hyperspectral data mining for a variety of purposes. In particular, it has been used as a preprocessing technique to reduce a very high dimensional data space coupled with its characteristically large volume to a manageable low dimensional space in which data analysis can be performed more efficiently. This study focuses on developing a workflow for hyperspectral image processing employing feature extraction using vegetation indices. Combined with feature selection to develop a DR by band selection (DRBS) method that searches for a subset of bands representing the original data, so that the information of interest in the data can be retained in the selected subset of bands. It is proposed to employ curve simplification techniques such as Douglas-Peucker to achieve this end. To perform the experiments we have used a hyperspectral image (HSI) taken by a drone flight with 0.31 m resolution, in a scenario that presents both vegetation and other architectural elements, with the presence of illuminated areas and shadows. Our results show that our method manages to reduce the amount of information by 96.4% of the HSI. In this case, preserving the most important features with a minimum level of loss, in most cases with mean square error very close to 0, this allows subsequently recreating the original data with high reliability.