Page 68 - 《高原气象》2026年第5期
P. 68

高     原      气     象                                 45 卷
              1284
                 2
             的 R 高达 0. 985, 而使用站点降水数据训练模型                       Hsu K L, Gao Xiaogang, Sorooshian S, et al., 1997. Precipitation es‐
                                                      2
             时, 模型的拟合程度较低, 训练集样本的 R 仅为                             timation from remotely sensed information using artificial neural
                                                                   networks[J].  Journal  of  Applied  Meteorology,  36(9):  1176-
             0. 452; 在样本值较低时, 回归模型的预测误差较
                                                                   1190.  DOI:  10.  1175/1520-0450(1997)036<1176:  pefrsi>2.  0.
             小, 但随着样本值增加, 预测误差也随之增大, 随
                                                                   co;2.
             机森林、 LightGBM 等算法在反演中也存在相同问                        Lazri M, Ameur S, 2016. A satellite rainfall retrieval technique over
             题, 这说明机器学习模型回归预测时存在系统性误                               northern  Algeria  based  on  the  probability  of  rainfall  intensities
             差, 本研究对极端降水的反演能力有待改进。                                 classification from MSG-SEVIRI[J]. Journal of Atmospheric and
                                                                   Solar-Terrestrial Physics, 147: 106-120. DOI: 10. 1016/j. jastp.
                 (3)  从降水反演模型得到的空间分布图来看,
                                                                   2016. 07. 015.
             模型在判断是否将降水及判断降水中心区域的位
                                                                Lazri M, Ameur S, Mohia Y, 2014. Instantaneous rainfall estimation
             置时, 具有良好的表现, 而在降水中心值的反演时                              using neural network from multispectral observations of SEVIRI
             误差较大且对于个别降水中心出现漏报现象; 从降                               radiometer and its application in estimation of daily and monthly
             水反演模型得到的单站降水来看, 在多个站点对降                               rainfall[J]. Advances in Space Research, 53(1): 138-155. DOI:
                                                                   10. 1016/j. asr. 2013. 10. 005.
             水峰值、 降水峰值时刻以及降水持续时间均比较准
                                                                Li Ziqi, 2022. Extracting spatial effects from machine learning model
             确, 但在个别站点也有对于降水峰值的错误判断以
                                                                   using local interpretation method: an example of SHAP and XG‐
             及在无降水时刻判断有降水发生。                                       Boost[J].  Computers,  Environment  and  Urban  Systems,  96:
                  在湿润半湿润半干旱气候背景下的黄河源区,                             101845. DOI: 10. 1016/j. compenvurbsys. 2022. 101845.
                                                        -1
             训练数据集中于 7 月强降水事件, >10. 0 mm·h 降                    Liu Nianqing, Ren Suling, Jiang Jianying, et al., 2023. AI-based esti‐
                                                                   mation  of  precipitation  in  the Tibetan  Plateau  using  multi-frame
             水样本占比不足 1%, 对非汛期降水及地形性降水
                                                                   FY-4A satellite data[J]. International Journal of Remote Sensing,
             的表征能力有限, 导致模型在弱降水事件中易出现
                                                                   44(20): 6523-6547. DOI: 10. 1080/01431161. 2023. 2270112.
             错报, 模型对强降水中心的反演存在系统性低估;                            Ouallouche F, Lazri M, Ameur S, 2018. Improvement of rainfall esti‐
             同时模型主要依赖统计关系, 对高原云降水微物理                               mation from MSG data using Random Forests classification and
             过程(如冰相粒子的辐射特性)的机理耦合不足。                                regression[J].  Atmospheric  Research,  211:  62-72.  DOI:  10.
                                                                   1016/j. atmosres. 2018. 05. 001.
                  未来的研究中, 可以尝试结合其他卫星的微波
                                                                Snoek J, Larochelle H, Adams R P, 2012. Practical Bayesian optimi‐
             波段数据与风云卫星红外波段数据, 构建多尺度的
                                                                   zation  of  machine  learning  algorithms[C]//Proceedings  of  the
             反演模型; 或在机器学习模型中引入云参数化方案                               26th  International  Conference  on  Neural  Information  Processing
             输出的微物理特征量作为约束或将反演模型与数                                 Systems-Volume 2. ACM: 2951-2959. DOI: 10. 5555/2999325.
             值预报模式耦合, 以体现降水的云物理过程。                                 2999464.
                                                                Thies B, Nauss T, Bendix J, 2008. Discriminating raining from non-
             参考文献(References):                                     raining clouds at mid-latitudes using meteosat second generation

                                                                   daytime  data[J].  Atmospheric  Chemistry  and  Physics,  8(9):
             Adler R F, Negri A J, 1988. A satellite infrared technique to estimate   2341-2349. DOI: 10. 5194/acp-8-2341-2008.
                 tropical convective and stratiform rainfall[J]. Journal of Applied   Wang  Gen,  Han Wei, Ye  Song,  et  al.,  2024.  FY-4A/AGRI  infrared
                 Meteorology, 27(1): 30-51. DOI: 10. 1175/1520-0450(1988)  brightness temperature estimation of precipitation based on multi-
                 027<0030: asitte>2. 0. co;2.                      model ensemble learning[J]. Earth and Space Science, 11(2):
             Arkin P A, 1979. The relationship between fractional coverage of high   e2023EA003311. DOI: 10. 1029/2023ea003311.
                 cloud  and  rainfall  accumulations  during  GATE  over  the  B-scale   陈利群, 刘昌明, 杨胜天, 等, 2006. 黄河源区降水遥感反演[J]. 中
                 array[J]. Monthly Weather Review, 107(10): 1382-1387. DOI:   国环境科学, 26(增刊 1): 87-91. Chen Liqun, Liu Changming,
                 10. 1175/1520-0493(1979)107<1382: trbfco>2. 0. co;2.   Yang Shengtian, et al., 2006. Reproduction of precipitation in the
             Chen  Tianqi,  Guestrin  C,  2016.  XGBoost:  a  scalable  tree  boosting   source regions of Yellow River with remote sensing[J]. China En‐
                 system[C]//Proceedings of the 22nd ACM SIGKDD International   vironmental  Science,  26(S1):  87-91.  DOI:  10.  3321/j.  issn:
                 Conference  on  Knowledge  Discovery  and  Data  Mining.  ACM:   1000-6923. 2006. z1. 022.
                 785-794. DOI: 10. 1145/2939672. 2939785.       陈亚玲, 文军, 杨传国, 等, 2024. 不同降水产品及WRF-Hydro模式
             Giannakos A, Feidas H, 2013. Classification of convective and strati‐  在黄河源区的适用性分析[J]. 大气科学, 48(3): 923-937.
                 form rain based on the spectral and textural features of Meteosat   Chen Yaling, Wen Jun, Yang Chuanguo, et al., 2024. Applicabil‐
                 Second Generation infrared data[J]. Theoretical and Applied Cli‐  ity of different precipitation products and the WRF-hydro model
                 matology, 113(3): 495-510. DOI: 10. 1007/s00704-012-0802-  over the source region of the Yellow River[J]. Chinese Journal of
                 z.                                                Atmospheric Sciences, 48(3): 923-937. DOI: 10. 3878/j. issn.
   63   64   65   66   67   68   69   70   71   72   73