Dimension Reduction of Combined Image and Elevation Remote Sensing Data Using UMAP, Autoencoders, and Variational Autoencoders: Investigation on Shaded Regions

Dimension reduction is a commonplace tool to visualize multi-dimensional data and reparametrize the features to have uniform, metric scales. With a concept of training a Machine Learning method with scarce training data in mind, we wish to investigate to what extent several well-known dimension reducers are suitable to separate very challenging remote sensing data, in particular, in shadow regions. The Potsdam data includes not only plenty of these regions, but also several seldom classes, such as vehicles or clutter. Hence, optical and elevation data as well as some additional features will be used as input for dimension reduction algorithms.