2022-06-20, 11:45–12:15, Room 2
How can the information content of large and complex remote sensing data sets be easily grasped and evaluated? And in which way is it possible to identify the potential of such data sets with respect to concrete objectives? Methods from the field of manifold learning, for which implementations are available as ready-to-use Python packages, are a good remedy. This talk focuses on the application of the dimension reduction algorithm Uniform Manifold Approximation and Projection (UMAP) for the visualization of high-dimensional remote sensing data.
In the evaluation and interpretation of remote sensing data, one often has to deal with high-dimensional data sets. The high number of dimensions of the observation space results, for example, from high spectral resolution in the case of hyperspectral data, the use of different polarizations and frequency bands in SAR systems, from temporal variation or from the combination of different sensor systems. The use of high-dimensional data offers opportunities and challenges at the same time. On the one hand, the use of many different measurements allows a better characterization of an observed scene. On the other hand, it is extremely difficult for human observers to get an overall view of all relevant information and to interpret their interaction. Furthermore, the high dimension of the observation space also complicates automatic pattern recognition, which is the basis of e.g. classification, regression or clustering. To counteract this problem, a number of dimension reduction methods have been developed, which aim at preserving relevant information in only a few components to the maximum possible extent while removing redundant information. One representative, which belongs to the field of manifold learning, is the algorithm UMAP presented in 2018 by McInnes et al..
In this talk, I will present applications of the UMAP algorithm for the visualization of high-dimensional remote sensing data. The focus is on the exploratory analysis of multi-frequency, full-polarimetric, interferometric SAR data. I will walk through the extraction of physical features that spans the high-dimensional observation space, the projection of the data in a 3-dimensional Euclidean space and the resulting opportunities for visual, comprehensive data analysis. In addition, I demonstrate how UMAP can be integrated as a pre-processing step for land cover classification, serving to simplify the training process and improve the explanatory power of classification results.