Extensive validation of a regularized neural-network technique for S-band polarimetric radar rainfall estimation

Raindrop Size Distribution (RSD) variability is one of the main factors affecting quantitative precipitation estimation from radar measurements. Dual-polarized radar systems enable the use of multi-parametric algorithms that generally improve the rainfall retrieval. Regression-based procedures are largely preferred by the operational community, simplicity being often considered synonymous of robustness. Despite the neural network (NN) capability to represent complex functions is well recognized in the estimation theory, its dissemination to the operational radar community is obstructed by an accompanying mystic halo. Rainfall retrieval problem is an ill-posed strongly non-linear problem. This means that the related inverse problem can be addressed only by resorting to the statistical analysis and by adding a priori information. Within this framework, the NN technique represents a powerful approach to design a retrieval algorithm in a more flexible and robust way than conventional methods such as linearized multivariate regression. Indeed, the selection of a NN topology, very often thought to be a “black box”, is theoretically equivalent to the choice of either a regression analytical model or a Bayesian probability model, whereas the NN training and test resembles the optimization step within the parameter estimation of statistical parametric relations.

An attempt to outline the potential benefit derived from the use of such NN approaches in radar rainfall estimation is carried out in the present work. A large radar data and surface gauge observation dataset collected in central Oklahoma during the multiyear Joint POlarization Experiment (JPOLE) field campaign is used to validate two neural network techniques: a) an ‘indirect' NN methodology based on the RSD retrieval and rainfall calculation; b) a ‘direct' NN methodology based on the rainfall retrieval. Both NN-based rainfall retrieval techniques are trained by a randomly-generated RSD dataset where independent RSD parameters are assumed within a climatological variability range. These assumptions ensure a broad applicability including the local expected correlation between the drop number concentration and mean diameter. Rainfall temporal accumulations from RSD retrieval-based methods are shown to be sensitive to the choice of a raindrop fall speed model. To minimize the impact of this choice, a further ‘direct' NN approach is tested. Proposed NN-based techniques exhibit bias and root mean square error characteristics comparable with those obtained from parametric relations, specifically optimized for the JPOLE dataset, indicating an appealing generalization capability with respect to the climatological context. All tested polarimetric relations are shown to be sensitive to hail contamination as inferred from the results of automatic polarimetric echo classification and available storm reports.