Sensitivity of Spatial Structure of Forecasts on Model Parameters

In recent work it was shown that Numerical Weather Prediction model parameters have a coherent and statistically significant effect on facets of forecasts that do not involve spatial information (e.g., domain average amount of precipitation). However, given the importance of spatial structure in forecast verification, it is important to assess the effect of model parameters on the spatial structure of forecasts. In this work various clustering methods are employed to quantify the spatial structure of gridded forecasts and observations, and the sensitivity of the resulting structures is assessed via multivariate (i.e. multiple response) regression methods. The clustering methods include Gaussian Mixture Models (GMM), Density-Based Spatial Clustering of Applications with Noise (DBSCAN), and other more traditional methods. The statistical significance of the results is performed by tests developed for multivariate linear models.