Accurate estimation of surface precipitation with high spatial and temporal resolution is critical for decision making regarding severe weather and water resources management. Polarimetric weather radar is the main operational instrument used for quantitative precipitation estimation (QPE). However, conventional parametric radar QPE algorithms such as the radar reflectivity (Z) and rain rate (R) relations cannot fully represent clouds and precipitation dynamics due to their dependency on local raindrop size distributions and the inherent parameterization errors. This article develops four deep learning (DL) models for polarimetric radar QPE (i.e., RQPENetD1, RQPENetD2, RQPENetV, RQPENetR) using different core building blocks. In particular, multi-dimensional polarimetric radar observations are utilized as input and surface gauge measurements are used as training labels. The feasibility and performance of these DL models are demonstrated and quantified using U.S. Weather Surveillance Radar - 1988 Doppler (WSR-88D) observations near Melbourne, Florida. The experimental results show that the dense blocks-based models (i.e., RQPENetD1 and RQPENetD2) have better performance than residual blocks, RepVGG blocks-based models (i.e., RQPENetR and RQPENetV) and five traditional Z-R relations. RQPENetD1 has the best quantitative performance scores, with a mean absolute error (MAE) of 1.58 mm, root mean squared error (RMSE) of 2.68 mm, normalized standard error (NSE) of 26%, and correlation of 0.92 for hourly rainfall estimates using independent rain gauge data as references. These results suggest that deep learning performs well in mapping the connection between polarimetric radar observations aloft and surface rainfall.
more »
« less
Improving Explainability of Deep Learning for Polarimetric Radar Rainfall Estimation
Machine learning‐based approaches demonstrate a significant potential in radar quantitative precipitation estimation (QPE) applications. In contrast to conventional methods that depend on local raindrop size distributions, deep learning (DL) can establish an effective mapping from three‐dimensional radar observations to ground rain rates. However, the lack of transparency in DL models poses challenges toward understanding the underlying physical mechanisms that drive their outcomes. This study aims to develop a DL‐based QPE system and provide a physical explanation of radar precipitation estimation process. This research is designed by employing a deep neural network consisting of two modules. The first module is a quantitative precipitation estimation network that has the capability to learn precipitation patterns and spatial distribution from multidimensional polarimetric radar observations. The second module introduces a quantitative precipitation estimation shapley additive explanations method to quantify the influence of each radar observable on the model estimate across various precipitation intensities.
more » « less- Award ID(s):
- 2239880
- NSF-PAR ID:
- 10512455
- Publisher / Repository:
- DOI PREFIX: 10.1029
- Date Published:
- Journal Name:
- Geophysical Research Letters
- Volume:
- 51
- Issue:
- 11
- ISSN:
- 0094-8276
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Real-time and accurate precipitation estimation is critical for environmental protection and water resources management. Compared to traditional methods, i.e., radar reflectivity (Z) and rainfall rate (R) relations, relying on local raindrop size distributions, the deep learning model can fit the functional relationship between radar observations and rainfall rate measurements. However, the black-box nature of deep learning models makes it difficult to explain the physical mechanisms behind their results. To address this problem, this study proposes DQPENet, a deep learning model for polarimetric radar QPE utilizing dense blocks. We employ a permutation test to understand the relative importance of different radar data input variables. Additionally, we propose a regression importance value (RIV) method for the precipitation estimation task to visualize feature importance regions. Our experimental results show that radar reflectivity and specific differential phase at the lowest elevation angle are the two most important observables for the model’s precipitation estimation. Furthermore, we find that radar data closer to the rain gauge are more influential on the model’s results, indicating that the deep learning model is able to capture the underlying physical mechanism of atmospheric data.more » « less
-
more » « less
Abstract We describe and evaluate adaptive conditional bias–penalized cokriging (CBPCK) for improved multisensor precipitation estimation using rain gauge data and remotely sensed quantitative precipitation estimates (QPE). The remotely sensed QPEs used are radar-only and radar–satellite-fused estimates. For comparative evaluation, true validation is carried out over the continental United States (CONUS) for 13–30 September 2015 and 7–9 October 2016. The hourly gauge data, radar-only QPE, and satellite QPE used are from the Hydrometeorological Automated Data System, Multi-Radar Multi-Sensor System, and Self-Calibrating Multivariate Precipitation Retrieval (SCaMPR), respectively. For radar–satellite fusion, conditional bias–penalized Fisher estimation is used. The reference merging technique compared is ordinary cokriging (OCK) used in the National Weather Service Multisensor Precipitation Estimator. It is shown that, beyond the reduction due to mean field bias (MFB) correction, both OCK and adaptive CBPCK additionally reduce the unconditional root-mean-square error (RMSE) of radar-only QPE by 9%–16% over the CONUS for the two periods, and that adaptive CBPCK is superior to OCK for estimation of hourly amounts exceeding 1 mm. When fused with the MFB-corrected radar QPE, the MFB-corrected SCaMPR QPE for September 2015 reduces the unconditional RMSE of the MFB-corrected radar by 4% and 6% over the entire and western half of the CONUS, respectively, but is inferior to the MFB-corrected radar for estimation of hourly amounts exceeding 7 mm. Adaptive CBPCK should hence be favored over OCK for estimation of significant amounts of precipitation despite larger computational cost, and the SCaMPR QPE should be used selectively in multisensor QPE.
-
Accurate precipitation retrieval using satellite sensors is still challenging due to the limitations on spatio-temporal sampling of the applied parametric retrieval algorithms. In this research, we propose a deep learning framework for precipitation retrieval using the observations from Advanced Baseline Imager (ABI), and Geostationary Lightning Mapper (GLM) on GOES-R satellite series. In particular, two deep Convolutional Neural Network (CNN) models are designed to detect and estimate the precipitation using the cloud-top brightness temperature from ABI and lightning flash rate from GLM. The precipitation estimates from the ground-based Multi-Radar/Multi-Sensor (MRMS) system are used as the target labels in the training phase. The experimental results show that in the testing phase, the proposed framework offers more accurate precipitation estimates than the current operational Rainfall Rate Quantitative Precipitation Estimate (RRQPE) product from GOES-R.more » « less
-
Missing or low-quality data regions usually happen to weather radars. One of the most common situations is beam blockage or partial beam blockage. Therefore, correction of weather radar observations that are partially or fully blocked is an indispensable step in radar data quality control and subsequent quantitative applications, especially in complex terrain environments such as the western United States. In this article, we propose a deep learning framework based on generative adversarial networks (GANs) for restoring partial beam blockage regions in polarimetric radar observations using local and global contextual information. Due to the diverse precipitation types, blockage conditions, and ground information in different areas, two radars deployed in two different regions characterized by different precipitation types are used to demonstrate the proposed methodology. Both are S-band operational Weather Surveillance Radar – 1988 Doppler (WSR-88D): KFWS located in Fort Worth, northern Texas, and KDAX located in Davis, northern California. For training the GAN model, this article simulates the partial beam blockage situations by manually cropping observation sectors of both KDAX and KFWS radar data. The trained models were tested using independent precipitation events in Texas and California to demonstrate the model effectiveness in inpainting “missing” data. In addition, this paper cross-tested the data with different precipitation features to examine the generalization capacity of the beam blockage correction models. The beam blockage correction performance is also compared with a traditional linear interpolation approach. The results show that for both domains the continuity of precipitation observations is greatly improved after applying the deep learning based inpainting approach. For the KFWS test data, some visible discrepancies exist between the results from models trained based on convective and stratiform precipitation events in Texas and California, respectively, yet both models outperform the traditional interpolation method. For the KDAX test data, both the model trained using the KFWS data from convective precipitation events in Texas and the model trained using KDAX data from stratiform precipitation events in California render a similar performance. Although ground truth is not available for the real blocked radar data, the repaired observations demonstrated a great potential for improved quantitative applications.more » « less
