Polarimetric radar-based rainfall estimation through adaptive learning with multi-source data from the NOAA meteorological assimilation data ingest system

Although polarimetric radar measurements are rich in information, traditional approaches of using them can only extract part of the information due to the limitations of the analytical tools. The performance of conventional radar rainfall estimation algorithms is highly dependent on the raindrop size distributions, which vary in different precipitation regions and/or different rainfall systems. It is challenging to remove the inherent parameterization error in "fixed" parametric radar rainfall relations. Recent studies have shown that deep learning techniques are effective in reducing such parameterization error and enhancing radar-based precipitation estimation. However, it is challenging to train a model that is applicable to a broad domain. Often, local rain gauge data would be required to retrain the model obtained for another domain. This study takes advantages of crowdsourced data from the NOAA Meteorological Assimilation Data Ingest System (MADIS), which is the national clearinghouse for weather and hydrologic observations that feed the NWS operational numerical weather and hydrologic prediction models. A convolutional neural network is utilized as benchmark, which incorporates a residual block to address the model degradation caused by the increased model depth. Through manipulation of the training process, the knowledge learned at one location is transferred to other domains characterized by different precipitation properties. The experimental results show that the proposed technique can improve precipitation estimation compared to conventional fixed-parameter rainfall algorithm.