| 引用本文: | 李克诚,陆建忠,张可睿,陆呈瑜,陈璞,袁明坤.基于改进U-Net的CMIP5全球气候模式降尺度方法及其在鄱阳湖流域的应用.湖泊科学,2022,34(1):320-333. DOI:10.18307/2022.0126 |
| Li Kecheng,Lu Jianzhong,Zhang Kerui,Lu Chengyu,Chen Pu,Yuan Mingkun.Spatial downscaling method and application of CMIP5 global climate models based on improved U-Net in Lake Poyang Basin. J. Lake Sci.2022,34(1):320-333. DOI:10.18307/2022.0126 |
|
| 摘要: |
| 利用降尺度方法对CMIP5全球气候模式进行空间降尺度并以此研究鄱阳湖流域未来气候时空变化趋势，能够为流域生态环境保护提供数据、技术和理论上的支持.通过简化原始网络结构，在网络首部添加插值层，采用反卷积算法作为上采样算法对传统U-Net网络进行改进，建立基于深度学习的气候模式空间降尺度模型（DLDM）.以1965-2005年鄱阳湖流域共18个气象站点的实测数据为基准，基于IPSL-CM5A-LR和BCC-CSM1.1两个气候模式，在拟合精度和流域极端气候事件模拟能力两方面对比验证了降尺度方法和降尺度后气候模式的模拟性能，结果表明基于DLDM的方法优于基于线性回归的传统统计降尺度方法，IPSL-CM5A-LR模式模拟效果优于BCC-CSM1.1模式.利用DLDM对RCP2.6和RCP8.5两个排放情景下的IPSL-CM5A-LR模式数据进行空间降尺度，基于2006-2100年流域高空间分辨率气候数据分析两情景下流域未来气候的时空变化特征，结果表明流域未来气温在两情景下均持续升高，表现出位于流域中北部、西部、东部和南部4个高温中心，RCP8.5情景下流域气温更高，升温趋势和局部周期变化更加明显；流域未来降水在两情景下先增加后减少，变化趋势较为平缓，表现出北部、中东部和南部3个降水中心，在未来呈现五次"枯-丰"交替，RCP8.5情景下流域降水更少且变化更为剧烈，在2075年左右出现突变并存在周期性振荡. |
| 关键词: CMIP5全球气候模式 改进U-Net网络 M-K分析 小波变换 鄱阳湖流域 |
| DOI：10.18307/2022.0126 |
| 分类号: |
| 基金项目:国家重点研发计划项目（2018YFC1506506）、武汉市应用基础前沿专项（2019020701011502）、江西省重点研发计划项目（20201BBG71002）、武汉大学大学生创新创业训练计划和测绘遥感信息工程国家重点实验室专项科研经费联合资助. |
|
| Spatial downscaling method and application of CMIP5 global climate models based on improved U-Net in Lake Poyang Basin |
|
Li Kecheng1, Lu Jianzhong2, Zhang Kerui1, Lu Chengyu1, Chen Pu1, Yuan Mingkun1
|
|
1.School of Resource and Environmental Sciences, Wuhan University, Wuhan 430079, P. R. China;2.State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan 430079, P. R. China
|
| Abstract: |
| Predicting spatiotemporal trend of the future climate in Lake Poyang Basin using CMIP5 global climate models downscaled by downscaling methods can provide data, technology, and theoretical support for the ecological and environmental protection of the basin. In current research area downscaling methods mainly include dynamic downscaling method and statistical downscaling method, among which statistical downscaling method based on linear regression is widely used thus defective in modeling non-linear relationships. In order to build up a better downscaling model, the deep learning ased spatial downscaling method (DLDM) which is more efficient in learning climate data features is established by simplifing the original network structure of U-Net, adding an interpolation layer at the head of the network, and using the deconvolution algorithm as upsampling algorithm. Then based on climate models as IPSL-CM5A-LR and BCC-CSM1.1, simulating ability of downscaled climate model is verified by making contrastive analysis to observed data from 18 meteorological stations in Lake Poyang Basin from 1965 to 2005 on fitting precision and ability to simulate extreme weather events. The results show that DLDM has higher fitting accuracy over the linear regression based statistical downscaling method and IPSL-CM5A-LR has higher fitting accuracy over BCC-CSM1.1. Based on high spatial resolution data of IPSL-CM5A-LR downscaled by DLDM from 2006 to 2100, the spatiotemporal trend of future climate in Lake Poyang Basin under RCP2.6 and RCP8.5 emissions scenarios is studied through M-K analysis and wavelet transform. The results show that the future temperature of the basin will continue to increase under the two scenarios, showing four high temperature centers located in the central north, west, east, and south of the basin. Under the RCP8.5 scenario, the temperature of the basin will be higher, and the warming trend and local cycle change will be more obvious. The future precipitation in the basin will increase first and then decrease under the two scenarios, and the changing trend is relatively gentle, showing three precipitation centers in the north, the central east, and the south of the basin. In the future, there will be "drought-wet" alternation for five times. Under the RCP8.5 scenario, the precipitation in the basin will be less and the change will be more intense, and there will be a sudden change and periodic oscillation at around 2075. |
| Key words: CMIP5 improved U-Net Mann-Kendall analysis wavelet transform Lake Poyang Basin |
