Self-Organizing Maps as a Dimension Reduction Approach for Spatial Global Sensitivity Analysis Visualization

Keywords: self-organizing maps, global sensitivity analysis, spatial models, geovisualization, uncertainty visualization
Session Type: Virtual Paper Abstract
Day: Monday
Session Start / End Time: 2/28/2022 03:40 PM (Eastern Time (US & Canada)) - 2/28/2022 05:00 PM (Eastern Time (US & Canada))
Room: Virtual 13

Abstract

Spatial global sensitivity analysis (SGSA) helps to understand how sensitive that our spatial models are to the uncertainty in the input layers, and then allows us to understand the significance of the input contribution. SGSA output is twofold: (1) first-order effects which are the results of the linear relation when every input layer or parameter varied one at a time, (2) high order effects where the nonlinear and complex interaction among input layers is depicted. Therefore, the result of SGSA is double proportional to the number of input layers. When there are n input layers, it will yield 2n sensitivity maps. This high dimensional input space becomes challenging to visualize, particularly considering that the limitations in human vision, the human cognitive system, and visual representations. One possible approach to tackle this voluminous visual load is to find similar patterns and project that similarity into a 2D surface by keeping the spatial characteristics. This paper presents the implementation of self-organizing maps (SOM), unsupervised learning networks that are used to tackle high-dimensional problems, as a dimension reduction approach for SGSA visualization. Once the winning neurons at the neural network are projected as the influence map, the SGSA is condensed into a 2D surface where only the most influential indices are encountered.