Gradient-based optimization to reduce uncertainty in radar rainfall estimates using deep learning techniques and in situ measurements from disdrometers

Traditionally, radar rainfall algorithms are derived through nonlinear regression of rain rates and simulated radar observables from raindrop size distribution (DSD) data. The performance of such empirical relations is highly dependent on the physical model of DSD and the parametric relation between the physical model and radar measurements. Often, the radar rainfall relations need to be adjusted for different climate regions, or different rainfall types (e.g., stratiform vs convective), or even in different precipitation regimes within a single storm system, in order to account for the space time variability in precipitation microphysics. The parameterization error inherent in the empirical relations is closely related to the rainfall intensity; it decreases as the rainfall intensity increases, and it will become rather stable when the rainfall intensity gets higher than about 100 mm hr^-1 (Bringi and Chandrasekar, 2001). Even if one can reduce all the random measurement errors in radar rainfall estimation through spatial and/or temporal averaging, the parameterization error is hard to eliminate. In this study, a deep learning (DL)-based approach is developed to estimate surface rainfall from ground radar measurements. This DL technique is able to extract the complex functional relation from high dimensional input space (i.e., radar data) to the target space (i.e., rainfall rate). Simulated polarimetric radar data based on DSD measurements in different climatological regimes are used to train and test the DL model. The trained model is also applied to real radar measurements to demonstrate its applicability. In addition, the dimensionality and representativeness of this DL model are investigated to resolve its error structure in different precipitation regimes. Preliminary results show that this DL model has superior performance to conventional fixed-parameter radar rainfall relations, especially when all the polarimetric observables are incorporated.