Work

Spatiotemporal Dual-Polarization Radar Quantitative Precipitation Estimation Using Convolutional Neural Networks

Year: 2026
Type: article
Abstract: Hourly quantitative precipitation estimation (QPE) from weather radar remains difficult during intense rainfall because fixed empirical relations and temporal averaging may not represent rapid scan-to-scan changes and localized echo structure. We evaluated five convolutional neural network configurations using S-band dual-polarization radar and rain-gauge observations in Guangzhou, China. A locally calibrated composite dual-polarization estimator (Composite-DP) provided a conventional reference. Each sample contains ZH, ZDR, and KDP from ten consecutive 0.5° scans over the preceding hour in a 3 × 10 × 21 × 21 tensor. Three-dimensional convolution spans time and two horizontal dimensions, without vertical-volume input. The dataset from 2018 to 2021 comprised 3,008,738 matched samples with rainfall ≥ 0.1 mm h−1, including 2,611,457 for training, 149,221 from six 2020 events for validation, and 141,428 from six 2021 events for testing. The models included a temporally averaged 2D baseline (2DPoly-QPENet), a time-resolved 3D baseline (3DPoly-QPENet), separate temporal (3DTime-QPENet) and multiscale (3DEcho-QPENet) extensions, and their fusion (3DHybrid-QPENet). The temporal extension represented the encoded ten-step sequence with a convolutional long short-term memory (ConvLSTM) layer, whereas the multiscale extension used an encoder–decoder pathway. On the independent test set, Composite-DP yielded an RMSE of 4.139 mm h−1. Compared with 2DPoly-QPENet, 3DPoly-QPENet had a lower mean absolute error (MAE; 1.321 versus 1.524 mm h−1) and root-mean-square error (RMSE; 3.088 versus 3.307 mm h−1), while its correlation coefficient (CC) was higher (0.860 versus 0.832). Among the individual extensions, 3DEcho-QPENet had lower overall errors than 3DTime-QPENet. Their fusion, 3DHybrid-QPENet, achieved the lowest overall MAE (1.270 mm h−1) and RMSE (2.971 mm h−1), with a Bias of −0.537 mm h−1 and the highest CC (0.866). Across the baseline and extension comparisons, the time-resolved 3D representation and the joint temporal and multiscale configuration were associated with lower hourly QPE errors. 3DHybrid-QPENet also had the lowest relative error (RE), MAE, and RMSE in all five rainfall-intensity classes. Above 50 mm h−1, however, its Bias remained −20.031 mm h−1, indicating persistent underestimation. Generalization beyond the study region remains to be evaluated.
Language: English
Open Access status: gold
Estimated APC paid: $3,258
OpenAlex ID: W7235700149
FWCI:
Cites: 24
Cited by:
Related to: 10