Wavelet-based spatial comparison technique for analysing and evaluating two-dimensional geophysical model fields

Abstract

Abstract. Complex numerical models of the Earth's environment, based around 3-D or 4-D time and space domains are routinely used for applications including climate predictions, weather forecasts, fishery management and environmental impact assessments. Quantitatively assessing the ability of these models to accurately reproduce geographical patterns at a range of spatial and temporal scales has always been a difficult problem to address. However, this is crucial if we are to rely on these models for decision making. Satellite data are potentially the only observational dataset able to cover the large spatial domains analysed by many types of geophysical models. Consequently optical wavelength satellite data is beginning to be used to evaluate model hindcast fields of terrestrial and marine environments. However, these satellite data invariably contain regions of occluded or missing data due to clouds, further complicating or impacting on any comparisons with the model. This work builds on a published methodology, that evaluates precipitation forecast using radar observations based on predefined absolute thresholds. It allows model skill to be evaluated at a range of spatial scales and rain intensities. Here we extend the original method to allow its generic application to a range of continuous and discontinuous geophysical data fields, and therefore allowing its use with optical satellite data. This is achieved through two major improvements to the original method: (i) all thresholds are determined based on the statistical distribution of the input data, so no a priori knowledge about the model fields being analysed is required and (ii) occluded data can be analysed without impacting on the metric results. The method can be used to assess a model's ability to simulate geographical patterns over a range of spatial scales. We illustrate how the method provides a compact and concise way of visualising the degree of agreement between spatial features in two datasets. The application of the new method, its handling of bias and occlusion and the advantages of the novel method are demonstrated through the analysis of model fields from a marine ecosystem model.