131 Non-parametric approaches to the modeling of orographic enhancement in radar images

Thursday, 29 September 2011
Grand Ballroom (William Penn Hotel)
Loris Foresti, University of Lausanne, Lausanne, Switzerland; and M. Kanevski and A. Pozdnoukhov

Handout (6.8 MB)

Reliable quantitative precipitation estimates and forecasting in Alpine terrains are complicated by the presence of orography which affects the quality of radar measurements on one side and is responsible of mechanisms of orographic precipitation in the lower levels of the atmosphere on the other side. Thus, it is necessary to analyze the causal relationship between the presence of particular orographic features and the corresponding precipitation enhancement.

This study aims at exploring a series of non-parametric data-driven classification algorithms for constructing maps of orographic enhancement susceptibility. The approach is based on the extraction of relevant topographic features from the digital elevation model at multiple spatial scales, mainly terrain convexity and flow dependent terrain slope, and on the detection of recurrent precipitation patterns in sequences of radar images such as stationary and persistent precipitation cells.

Non-parametric support vector machine (SVM) classifiers are applied to partition the set of precipitation cells into orographic and non-orographic ones in the high-dimensional space of topographic and flow-related features. The probabilistic interpretation of SVM depicts the membership to the class composed of orographic cells and can be interpreted as a likelihood of orographic enhancement.

Radar-based rainfall estimates from the network of three Swiss C-band radars are used to demonstrate the applicability of the methodology and to test a series of case studies including the widespread precipitation event of August 2005 in the Alps. The main questions which are explored are the investigation of the most suitable SVM models for orographic enhancement mapping, the optimal sampling of precipitation cells to avoid overfitting of the data as well as the exploration of semi-parametric approaches to account for prior meteorological knowledge.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner