Adaptive moment estimation for polarimetric weather radar using explainable deep learning-based estimators

Abstract

This paper presents machine learning-based approaches to improve moment estimation for polarimetrie weather radar. A novel weighted multilag estimator (WMLE) is proposed, with adaptively learned weights optimized using deep learning techniques. Two approaches of multilayer perceptron (MLP) and convolutional neural network (CNN) are used to implement WMLE. The performance of WMLE is evaluated using the measurements from the Next-Generation Weather Radar (NEXRAD) system. Experimental results demonstrate that the WMLE significantly improves polarimetric data quality, achieving lower root mean square error and standard deviation compared to conventional 0-Lag and 1-Lag estimators. In addition, the CNN-based estimator surpasses its MLP counterpart by leveraging spatial information in the input data and producing content-aware dynamic adaptive weights. Furthermore, the CNN-based estimator achieves superior radar data quality using data from only 32 pulses, compared with the 0-Lag and 1-Lag estimators using 64 pulses. Moreover, the CNN model demonstrates physical explainability, as its learned weights exhibit meaningful correlations with the characteristics of NEXRAD data.