Copula Method and Neural Networks for X-Band Polarimetric Radar Rainfall Retrieval in West Africa

In the context of climate change, countries in West Africa are faced with recurrent flooding with catastrophic consequences, that makes it imperative to have access to rainfall information on fine spatial and temporal scales for better monitoring and prediction of these phenomena, as could be provided by weather radars. Based on an extensive archive of data from the X-band polarimetric radar and rain gauges observations gathered during the intensive AMMA campaigns in 2006–2007 and the Megha-Tropiques satellite measurement validation programme in 2010 in West Africa, we (i) simulated jointly realistic data for polarimetric radar variables and rain intensity using copula, and (ii) assessed rain rate estimation methods based on neural network (NN) inversion techniques and non-linearly calibrated parametric algorithms. The assessment of rainfall rate retrieval by these estimators is carried out using the part of the observations database not employed for calibration steps. The multiparametric algorithms R(ZH,KDP) and R(ZDR,KDP) perform better than R(ZH,ZDR) and R(ZH,ZDR,KDP), especially since they are calibrated using copulas with upper tail dependencies, with KGE ranging in 0.68-0.75 and 0.79-0.82, respectively versus ranges of 0.40-0.64 and 0.20–0.51, for the two latter estimators. The neural network-based estimators RNN(ZDR,KDP) and RNN(ZH,KDP), show KGE score characteristics comparable to those obtained from the best parametric relations, specifically optimized for the synthetic copula-based dataset. However, the neural network-based estimators were shown to be more robust when applied to a specific rainfall event. More specifically, neural network-based estimators trained on synthetic data are sensitive to the copulas' ability to capture the dependence between the variables of interest over the entire distribution of joint values. This leads to a near-cancellation of sensitivity to variability in the raindrop size distribution, as shown the coefficients of correlation near 1, especially for RNN(ZDR,KDP), and for less extent RNN(ZH,KDP).