Interpretable Deep Learning for Polarimetric Radar Rainfall Estimation

Polarimetric radar-based rainfall estimates are typically derived through empirical parametric relations obtained from nonlinear regression between measured rain rates and simulated radar observables from disdrometer data. The performance of such empirical relations is highly dependent on the sample size and representativeness of the measured raindrop size distribution, which vary in different precipitation regimes or even within a single storm system. Even if one can eliminate all the random errors associated with polarimetric radar measurements, the parameterization error inherent in the empirical relations used to estimate rainfall rate from polarimetric radar measurements is hard to reduce. In addition, it is difficult to estimate surface rain rates with the radar measurements aloft, especially during the precipitation events characterized by varying precipitation microphysics with height. Recent studies have shown that artificial intelligence (AI) is effective in reducing these parameterization errors and improving the accuracy of radar-based precipitation estimation. However, it is challenging to train a deep learning AI model that is applicable to a variety of locations or precipitation regimes. Often, local rain gauge data would be required to retrain the model developed in a domain with different precipitation characteristics. To this end, this research uses a deep convolutional neural network (CNN) as benchmark to design a transfer learning framework to incorporate the knowledge learned at one location to the other locations which feature different precipitation characteristics. Saliency maps are used to interpret the underlying model physics and quantify the impacts of input features on the performance of precipitation estimation. Extensive experiments are performed to explain the transfer learning model by investigating the saliency maps on the activation of a specific neuron and different groups of neurons. The results demonstrate that this explainable machine learning framework can significantly improve precipitation estimation accuracy compared to conventional fixed-parameter rainfall algorithms.