Visualization of Multi-dimensional MISR Datasets Using Self-Organizing Map

Many techniques exist for visualization of high dimensional datasets including Parallel Coordinates, Projection Pursuit, and Self-Organizing Map (SOM), but none of these are particularly well suited to satellite data. Remote sensing datasets are typically highly multivariate, but also have spatial structure. In analyzing such data, it is critical to maintain the spatial context within which multivariate relationships exist. Only then can we begin to investigate how those relationships change spatially, and connect observed phenomena to physical processes that may explain them. We present an analysis and visualization system called SOM_VIS that applies an enhanced SOM algorithm proposed by Todd & Kirby [1] to multi-dimensional image datasets in a way that maintains spatial context. We first use SOM to project high-dimensional data into a non-uniform 3D lattice structure. The lattice structure is then mapped to a color space to serve as a colormap for the image. The Voronoi cell refinement algorithm is then used to map the SOM lattice structure to various levels of color resolution. The final result is a false color image with similar colors representing similar characteristics across all its data dimensions. We demonstrate this system using data from JPL's Multi-angle Imaging Spectro-Radiometer (MISR), which looks at Earth and its atmosphere in 36 channels: all combinations of four spectral bands and nine view angles. The SOM_VIS tool consists of a data control panel for users to select a subset from MISR's Level 1B Radiance data products, and a training control panel for users to choose various parameters for SOM training. These include the size of the SOM lattice, the method used to modify the control vectors towards the input training vector, convergence rate, and number of Voronoi regions. Also, the SOM_VIS system contains a multi-window display system allowing users to view false color SOM images and the corresponding color maps for trained SOM lattices. In addition, SOM_VIS allows users to interactively select an input vector, relate it to its corresponding SOM vector or location in color space, and plot it in 36 dimensions along with its corresponding SOM vector and Euclidean distances between the two. In this presentation, we report results of our exploration of selected MISR datasets using SOM_VIS. Early experience shows that SOM_VIS is not only able to extract features common to channels, but also can identify subtle differences among channels or signals only visible in a few channels. We also present some quantitative performance measures comparing the SOM algorithm to other traditional clustering algorithms. [1] A. Todd & M. Kirby, "Data Visualization via Structured Voronoi Cell Refinement", SIAM Workshop on Mining Scientific Datasets, p.45-52, April 2001