Self-organizing maps as a dimension reduction approach for spatial global sensitivity analysis visualization

Abstract

Spatial global sensitivity analysis (SGSA) reveals and ranks the input–output relation in spatial models. The SGSA output is twofold: (1) first-order effects which are the linear relations of every input layer with the output; and (2) high-order effects where the nonlinear interaction among input layers is depicted. The resulting sensitivity maps are twice the number of input layers which is challenging to visualize, considering the limitations of the human cognitive system or visual representations. Finding similar patterns and projecting that similarity into a 2D surface will help to tackle this voluminous visual load. This article presents the implementation of self-organizing maps (SOM), a type of artificial neural network, as a dimension reduction approach for SGSA visualization. SOM is also used for feature selection to identify the most relevant feature for model uncertainty. The winning neurons at SOM are projected as the influence map and the results are compared with conventional visualization techniques.