Page 96 - 《高原气象》2026年第3期
P. 96

高     原      气     象                                 45 卷
              704





                            CBAM U-Net Model-Based Precipitation Estimation for

                                       Chengdu X-band Phased Array Radar


                                         1
                                                             2
                               ZHOU Cong , ZHANG Chenghong , DONG Yuanchang , ZHANG Tao     1
                                                                               2
                                   (1. Chengdu Meteorological Observatory, Chengdu  611134, Sichuan, China;
                         2. Institute of Plateau Meteorology, China Meteorological Administration, Chengdu  610218, Sichuan, China)

             Abstract: The frequent occurrence of heavy precipitation events poses severe challenges to urban disaster pre‐
             vention and mitigation.  High-resolution quantitative precipitation estimation (QPE) products are of great impor‐
             tance for monitoring intense rainfall processes, while X-band phased array radars (X-PAR), delivering detection
             data with high spatiotemporal resolution, provide a robust foundation for advanced QPE research.  Due to the
             limitations of traditional radar-based QPE methods, such as insufficient nonlinear mapping capabilities and the
             lack of dynamic adaptive learning mechanisms, a minute-scale rainfall QPE method based on the CBAM U-Net
             model is proposed.  This model enhances the learning capability for key precipitation regions and feature chan‐
             nels by incorporating the Channel and Spatial Attention Module (CBAM) and optimizes the temporal alignment
             between  radar  and  ground  rain  gauge  observations  by  introducing  a  temporal  dimension  to  the  model  input.
             X-PAR data from four radars in Chengdu’s plain area during July-September 2023 -2024 were used to construct
             training and testing datasets.  Additionally, the training process was optimized through the integration of learning
             rate warm-up and cosine annealing strategies, which accelerated the convergence speed and improved the model’s
             robustness.  Evaluation metrics, including regression indices (CC, R MSE , R MAE , R ) and classification metrics
                                                                                     MB
             (POD, FAR, CSI), were applied to compare the performance of the CBAM U-Net against Attention U-Net,
             U-Net, and two Z-R methods.  Results show that deep learning models significantly outperform traditional Z-R
             methods in both error metrics and stability.  Among the three models, CBAM U-Net achieves the best perfor‐
                                                   -1
             mance (CC=0. 665,  R MSE =0. 331  mm·min ,  R MAE =44. 651%,  R =7. 324%),  followed  by Attention  U-Net,
                                                                      MB
             while  U-Net  shows  significant  underestimation.   In  multi-threshold  classification  evaluations,  CBAM  U-Net
             achieves  CSI  of  0. 655  and  0. 485  at  thresholds  of  0. 34  mm·min (corresponding  to  20  mm·h )  and
                                                                           -1
                                                                                                       -1
             0. 67 mm·min⁻¹ (corresponding to 40 mm·h ) respectively, which are notably higher than those of the Atten‐
                                                     -1
             tion U-Net (0. 609, 0. 408) and U-Net (0. 537, 0. 160).  Case studies further verify that CBAM U-Net has a
             stronger  capability  to  capture  temporal  features.   During  the  "pulse-type"  precipitation  event  on  10  September
             2024, CBAM U-Net accurately identifies the peaks and troughs of precipitation curves and effectively character‐
             izes the temporal variation features of precipitation, performing better than U-Net and Attention U-Net.  During
             the stable precipitation process on 29 September 2024, CBAM U-Net successfully captures the overall trend of
             precipitation  changes  but  shows  insufficient  sensitivity  to  minor  fluctuations  within  the  range  of  0. 1~0. 2
             mm·min .  The CBAM module effectively improves the model's capability to capture temporal characteristics of
                     -1
             heavy precipitation via its dual attention mechanism.  However, due to the inherent systematic smoothing effect
             of deep learning models, notable underestimation still occurs for extreme heavy precipitation events with minute
                                      -1
             rainfall rates ≥1. 17 mm·min .
             Key words: quantitative precipitation estimation; X-band phased array radar; short-time heavy rainfall; CBAM
             U-Net
   91   92   93   94   95   96   97   98   99   100   101