Improving Interpretation of A Deep Learning Model for Radar Rainfall Mapping

Machine learning-based approaches demonstrate significant potential for use in quantitative precipitation estimation (QPE) applications. In contrast with conventional parametric methods that depend on local raindrop size distributions, such as radar reflectivity ($Z$) and rainfall rate ($R$) relationships, a deep learning (DL) model can establish a functional approximation between three-dimensional radar observations and ground rain rate measurements. However, the lack of transparency in deep learning (DL) models poses challenges toward understanding the underlying physical mechanisms that drive their outcomes. This obstacle hampers the practical implementation of DL-based QPE networks. To address this issue, the current study aims to develop a QPE system for polarimetric radar that offers a comprehensible explanation of its functions. This system is created by employing a deep neural network consisting of two main modules. The first module is designed to introduce a QPENet that is specifically tailored for polarimetric radar. The study utilizes dual-polarization radar observations and rain gauge measurements as inputs and labels, respectively. The second module includes a QPESHAP explainability method for elucidating the factors that influence the model precipitation estimates across varying precipitation levels. Moreover, the study offers more accurate visual explanations by visualizing the interactions among the input radar variables. The experimental results showcase the favorable performance of QPENet across different rainfall intensities, with each observable result exhibiting distinct effects in light, moderate, and heavy precipitation scenarios. Furthermore, the findings derived from the QPESHAP analysis suggest that the deep learning model effectively captures the fundamental physical mechanisms associated with precipitation data.