Correction of Monthly Average 2 m Temperature Prediction: A Method Based
-
摘要: 作为减少短期气候预测误差的技术，数据订正成为了重要的研究方向。而深度学习作为一种新兴方法已经应用到数据订正技术中，其中常用的模型是U-Net，但它存在不可避免的缺陷。第一，U-Net基于卷积神经网络，但是受限于卷积神经网络的小感受野，这导致U-Net不能从全局的角度学习空间特征；第二，U-Net的下采样操作容易丢失图像细节信息。这两点都影响了该模型的订正性能。因此采取以下两个措施进行改进，一是将原模型与能够学习图片全局特征的Vision Transformer有机结合起来，使其能够从全局的角度学习空间特征；二是引入UNet 3+模型中的全尺度连接操作，弥补原下采样中丢失的图像细节信息。改进之后的模型称为UNet-Former 3+，在CMIP6中月平均2 m气温的春季和冬季数据集上进行订正实验，ERA5为实验标签。模型会与分位数映射、岭回归、U-Net、CU-Net、Dense-CUnet和RA-UNet这六种订正方法进行对比。实验结果表明，此模型在两个季节的平均绝对误差都下降49%，均方根误差都下降57%，两者都低于上述六种方法。总之，UNet-Former 3+在春季和冬季的订正效果优于上述六种方法。
-
关键词:
- 短期气候预测 /
- 数据订正 /
- Vision Transformer /
- 全尺度连接 /
- UNet-Former 3+
Abstract: As a technique to reduce the error in short-term climate prediction, bias correction has become an important research direction. This study explored the application of deep learning techniques in bias correction, focusing on the U-Net model, which, despite its popularity, has inherent limitations. First, UNet is based on a convolutional neural network, which has a limited receptive field, preventing it from fully capturing spatial features from a global perspective. Second, the subsampling operation in U-Net often leads to a loss of important image details. To address these issues, we implemented the following two measures. First, we integrated the original model with a Vision Transformer, which is capable of learning global features, thereby overcoming the limitation of the convolutional neural network's small receptive field. The second was to introduce the full-scale connection operation from the UNet 3+ model and compensate for the image details lost in the original down-sampling process in the decoder. The improved model is called UNet-Former 3 +. A correction experiment was carried out on the spring and winter datasets of the monthly average 2m temperature in CMIP6, with ERA5 as the experimental label. We compared its performance against six other correction methods: quantile mapping, ridge regression, UNet, CU-Net, Dense-CUnet, and RA-UNet. The experimental results show that the average absolute error and root mean square error of this model were reduced by 49% and 57%, respectively, outperforming the other six methods. Overall, the UNet-Former 3+ model demonstrated superior correction performance for both spring and winter seasons. -
图 6 MAE空间分布图（类别从左到右依次为CMIP6、分位数映射、岭回归、U-Net、CU-Net、Dense-CUnet、RA-UNet和UNet-Former 3+）
图 10 2022年1月2 m气温空间分布图（类别从上到下、从左到右依次为CMIP6、分位数映射、岭回归、U-Net、CU-Net、Dense-CUnet、RA-UNet、UNet-Former 3+和ERA5）
表 1 MAE和RMSE得分（单位：K）
季节 方法 MAE RMSE 春 CMIP6 3.782 4 5.374 分位数映射 2.819 6 4.229 3 岭回归 2.792 5 4.101 4 U-Net 2.451 4 3.530 7 CU-Net 2.317 0 3.255 3 Dense-CUnet 2.476 0 3.667 5 RA-UNet 2.296 5 3.260 2 UNet-Former 3+ 1.931 7 2.745 8 冬 CMIP6 3.936 1 5.515 5 分位数映射 3.769 1 5.804 7 岭回归 3.521 8 5.301 4 U-Net 2.766 7 3.770 4 CU-Net 2.786 4 3.612 5 Dense-CUnet 2.701 0 3.900 7 RA-UNet 2.516 4 3.439 7 UNet-Former 3+ 1.686 9 2.386 5 
下载: 导出CSV
表 2 春、冬季消融实验的结果（单位：K）
季节 ViT 全尺度连接 MAE RMSE 春 2.451 4 3.530 7 ✔ 2.068 1 2.856 1 ✔ ✔ 1.931 7 2.745 8 冬 2.766 7 3.770 4 ✔ 1.780 6 2.483 5 ✔ ✔ 1.686 9 2.386 5
下载: 导出CSV
-
[1] 高雪. 浅谈我国短期气候预测的现状及改革措施[J]. 低碳世界, 2019, 9(8): 313-314. [2] 孙照渤, 陈海山. 短期气候预测: 南京信息工程大学60年回顾与展望[J]. 大气科学学报, 2020, 43(5): 745-767. [3] 杨成荫, 赵苏璇, 程立国, 等. 短期区域气候预测系统及回报试验[J]. 解放军理工大学学报, 2016, 17(3): 289-295. [4] 曲美慧, 涂钢, 刘洋, 等. JL-NZCv1.0短期气候预测系统对吉林省春、夏季气温、降水预测的初步检验评估[J]. 气象灾害防御, 2022, 29(1): 16-21. [5] 陈明轩, 张延彪, 韩雷, 等. 数值天气预报多要素深度学习融合订正方法[J]. 气象学报, 2022, 80(1): 153-167. [6] 陆虹, 何慧, 覃卫坚, 等. 广西短期气候预测业务和技术发展回顾与展望[J]. 气象研究与应用, 2020, 41(4): 48-55. [7] 任宏利, 张培群, 李维京, 等. 提高月预报业务水平的动力相似集合方法[J]. 气象学报, 2014, 72(4): 723-730. [8] HUANG Y Y, WANG H J. A possible approach for decadal prediction of the PDO[J]. J Meteor Res, 2020, 34(1): 63-72. [9] HUANG Y Y, WANG H J. Is the regional precipitation predictable in decadal scale? A possible approach for the decadal prediction of the summer precipitation over North China[J]. Earth and Space Science, 2020, 7(1): 986-996. [10] 曲美慧, 涂钢, 李尚锋, 等. NZC-PCCSM4吉林省本地化及其在2019年春季气候预测中的应用[J]. 气象灾害防御, 2020, 27(3): 32-37. [11] 林广洪, 朱碧莹, 陈杰, 等. 多气候模式的全国月降水预测能力评价及偏差校正[J]. 中国农村水利水电, 2023(6): 47-56, 65. [12] 邱学兴, 王东勇, 陈宝峰. T639模式预报系统误差统计和订正方法研究[J]. 气象, 2012, 38(5): 526-532. [13] 毛波, 杨昊, 周世杰, 等. 基于CMA-REPS格点预报数据的深度学习风速订正方法[J]. 成都信息工程大学学报, 2023, 38(3): 264-270. [14] 方巍, 庞林, 王楠, 等. 人工智能在短临降水预报中应用研究综述[J]. 南京信息工程大学学报, 2020, 12(4): 406-420. [15] 杨淑贤, 零丰华, 应武杉, 等. 人工智能技术气候预测应用简介[J]. 大气科学学报, 2022, 45(5): 641-659. [16] HU Y F, YIN F K, ZHANG W M. Deep learning-based precipitation bias correction approach for Yin-He global spectral model[J]. Meteor Appl, 2021, 28(5): 2 032-2 045. [17] HESS P, NIKLAS B. Deep learning for improving numerical weather prediction of heavy rainfall[J]. Journal of Advances in Modeling Earth Systems, 2022, 14(3): 2 765-2 775. [18] XIANG T, ZHI X F, GUO W J, et al. Ten-meter wind speed forecast correction in southwest China based on U-Net neural network[J]. Atmosphere, 2023, 14(9): 1 355-1 376. [19] CHEN M X, HAN L, CHEN K K, et al. A deep learning method for bias correction of ECMWF 24-240 h forecasts[J]. Adv Atmos Sci, 2021, 38(9): 1 444-1 459. [20] FEI T H, HUANG B H, WANG X, et al. A hybrid deep learning model for the bias correction of SST numerical forecast products using satellite data[J]. Remote Sensing, 2022, 14(6): 1 339-1 357. [21] HAN K, WANG Y, CHEN H, et al. A survey on vision transformer[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 45(1): 87-110. [22] VALANARASU J M J, SINDAGI V A, HACIHALILOGLU I, et al. Kiu-net: Overcomplete convolutional architectures for biomedical image and volumetric segmentation[J]. IEEE Transactions on Medical Imaging, 2021, 41(4): 965-976. [23] 周嘉月, 卢麾, 阳坤, 等. 基于CMIP6的中高温升情景对中国未来径流的预估[J]. 中国科学: 地球科学, 2023, 53(03): 505-524. [24] WEI X K, WANG G J, FENG D H, et al. Deep-learning-based harmonization and super-resolution of near-surface air temperature from CMIP6 models (1850—2100)[J]. Int J Climatol, 2023, 43(03): 1 461-1 479. [25] 刘婷婷, 朱秀芳, 张世喆, 等. ERA5再分析地面气温数据在中国区域的适用性分析[J]. 热带气象学报, 2023, 39(1): 78-88. [26] UNAL A, ASAN B, SEZEN I, et al. Climate model-driven seasonal forecasting approach with deep learning[J]. Environmental Data Science, 2023, 2: e29. -
点击查看大图
图(10) / 表(2)
计量
- 文章访问数: 13
- HTML全文浏览量: 5
- PDF下载量: 1
- 被引次数: 0
粤公网安备 4401069904700003号