Using K-Means Clustering for a Spatial Analysis of Multivariate and Time-Varying Microclimate Data

Abstract

In this study, we propose a k-means clustering algorithm combined with glyph-based encoding method to analyze the spatial distribution and dependence of multivariate, time-varying 3D microclimate data. We obtained five climate variables, i.e. air and surface temperature, specific humidity, direct shortwave radiation and sensible heat flux, from an ENVI-met R simulation of a residential neighborhood in Phoenix, AZ. In a preprocessing step, we aggregated the 3D gridded simulation data by adding up value differences between two consecutive time steps for each grid cell over the entire simulation time to get a highly compressed view of the data without losing the spatial context. K-means clustering was then conducted in coordinate space by weighting each grid cell based on its difference to the spatial mean of temporal value differences. To reduce occlusion and to encode additional cluster member information, the visualization focused on the k-means cluster centroids. Resulting images show that the applied technique is suitable to provide a first insight into the spatial relationship of features based on their temporal variability.