一种降水降尺度模型:地形约束下的条件生成对抗网络  

A Precipitation Downscaling Model:Conditional Generative Adversarial Networks under Terrain Constraints

作　　者：杜晓婉 陈曦[1,2] 郑宏伟[1,2] 刘英[1,2] 刘铁[1,2] 包安明[1,2] 胡萍[4] DU Xiaowan;CHEN Xi;ZHENG Hongwei;LIU Ying;LIU Tie;BAO Anming;HU Ping(State Key Laboratory of Desert and Oasis Ecology,Xinjiang Institute of Ecology and Geography,Chinese Academy of Sciences,Urumqi 830011,China;Research Center for Ecology and Environment of Central Asia,Chinese Academy of Sciences,Urumqi 830011,China;College of Resources and Environment,University of Chinese Academy of Sciences,Beijing 100049,China;School of Information Science and Engineering,Xinjiang University,Urumqi 830017,China)

机构地区：[1]中国科学院新疆生态与地理研究所荒漠与绿洲生态国家重点实验室,乌鲁木齐830011 [2]中国科学院中亚生态与环境研究中心,乌鲁木齐830011 [3]中国科学院大学,北京100049 [4]新疆大学信息科学与工程学院,乌鲁木齐830017

出　　处：《地球信息科学学报》2023年第8期1586-1600,共15页Journal of Geo-information Science

基　　金：国家自然科学基金重点基金项目(42230708);国家自然科学基金面上项目(42071245);中国科学院先导研究计划项目(XDA20060303);新疆维吾尔自治区重点研发专项科技计划项目(2022B03001-3)。

摘　　要：中亚降水数据存在缺失、地理偏差、分辨率低和采集难度大等问题。近年来,神经网络模型被广泛应用于降水降尺度研究。然而,由于山区自然环境复杂多变,普通神经网络模型的预测结果难以解释且适用性差。为此,本文以地理差异分析作为先验知识约束生成式对抗网络,构建一种新的降水降尺度模型,提高了阿姆河流域复杂环境下降水数据的空间分辨率和精度。首先,依据地形数据通过空间变形模型对输入的Climate Research Units Time Series(CRUTS)降水数据进行空间校正。然后,输入校正后的CRUTS降水数据、气温风速湿度等同化数据及遥感数据到条件生成式对抗网络,重建高分辨率降水数据。最后,考虑到山区降水的各向异性,尤其在地形复杂的上游区域,该模型基于气象站点的真值,对降水数据进行了反距离权重的地理差异分析。结果表明,基于地理差异约束生成式对抗网络的降水降尺度模型能够提升复杂环境降水数据的分辨率和精度。针对中亚阿姆河流域的实验表明,本方法可将CRUTS降水数据的分辨率由55 km提升至11 km,其R2值增加了0.34,均方根误差(RMSE)和平均绝对误差(MAE)分别减小19.4 mm和10.65 mm,偏差(Bias)也由原来的0.24降至0.08。本文为数据采集难、地形地貌复杂区域的降水数据空间分辨率的提高,提供了鲁棒性好、普适性强的方法和思路。Most of the precipitation datasets in Central Asia have problems such as data missing,geographical bias and outliers,low resolution,and so on.The normal prediction results obtained by most machine learning methods are usually hard to interpret,not only due to the uncertainties from input information but also due to the complicated global geographical environments as well as the underlying local geographical conditions.In this paper,to overcome this problem,we proposed a novel downscaling precipitation model to adjust and optimize the precipitation computation results from Conditional Generative Adversarial Networks(CGAN)using an inverse distance weighting method based on the prior information of geographical differences of local digital terrain model and multiple weather stations.In this study,the Amu Darya River Basin was selected as the research area due to its various geographical environment and complicated topographic and geographical conditions.First,the input Climate Research Units(CRU)precipitation data with 55 km resolution were spatially corrected based on the topographic map using the spatial deformation model.The spatial deformation model was extended from spatial transformation network methods.Second,we input the corrected CRU precipitation data,temperature,wind speed,humidity equivalent data,and remote sensing data to the CGAN computing framework for high-resolution precipitation reconstruction.The experiment adopted the cross-validation method,taking 80%of the data as the training set,and the remaining 20%as the verification set.The test set contained 20 raster maps of annual precipitation from 2000 to 2019.The model was built based on pytorch 1.10.0,the batch size was 16,and the learning rate was 0.0003.The epoch was 8000 iterations in the Adam optimizer for gradient descent.Finally,the precipitation data of meteorological stations were used as the true values for analyzing the geographical differences of inverse distance weights and the accuracy of the corrected precipitation grid data.The result

关 键 词：降尺度 降水 条件生成式对抗网络 空间变形模型 地理差异分析 反距离权重 图像高分辨率重建 地形约束 

分 类 号：P426.6[天文地球—大气科学及气象学] P208