基于数据融合和偏差校正的中国大陆降水氧稳定同位素数据集（1870-2017）
收藏资源简介：
数据集包括1870-2017年中国大陆地区的降水氧稳定同位素，空间分辨率为50-60 km，时间分辨率为月。为了充分利用观测数据，并整合各引入同位素的大气环流模式（iGCM）的优势，采用了数据融合与偏差校正相结合的方法。数据融合中引入了基于物理的辅助数据，包括高程和气象数据（SRTM DEM，国家气象信息中心 中国地面气温和降水月值0.5°格点数据），以丰富融合过程中的气候和地形信息。数据集各时段的生成方法具体为：（1）1870-1957年，使用线性缩放（LS）和分位数匹配（DT）两种偏差校正方法对HadAM3模拟序列进行校正，数据集使用两种校正结果的集合平均；（2）1958-1968年，使用LS和DT两种偏差校正方法对CAM2和HadAM3两个模拟序列进行校正，数据集使用两个序列两种校正结果的集合平均；（3）1969-1978年，使用卷积神经网络（CNN）融合方法对三个模式的四个模拟序列（CAM2，GISS E free，GISS E nudged和HadAM3）和辅助数据进行融合；（4）1979-2001年，使用CNN融合方法对六个模式的九个模拟序列（CAM2，GISS E free，GISS E nudged，HadAM3，IsoGSM2，LMDZ4 free，LMDZ4 nudged，LMDZ4 zoomed和MIROC32）和辅助数据进行融合；（5）2002-2007年，使用CNN融合方法对四个模式的七个模拟序列（GISS E free，GISS E nudged，IsoGSM2，LMDZ4 free，LMDZ4 nudged，LMDZ4 zoomed和MIROC32）和辅助数据进行融合；（6）2008-2017年，使用LS和DT两种偏差校正方法对IsoGSM2和LMDZ4 zoomed两个模拟序列进行校正，数据集使用两个序列两种校正结果的集合平均。 在构建数据集之前，比较了三种数据融合方法和两种偏差校正方法的性能。结果表明，基于CNN的融合方法表现最好（与观测数据的相关系数大于0.95，均方根误差小于1‰），误差反向传播神经网络和长短期记忆人工神经网络的表现相近，且优于偏差校正方法。因此，使用CNN融合方法生成1969-2007年iGCM共同时段的数据集，并使用偏差校正方法生成其余时段的数据集。与 iGCM 模拟相比，构建的数据集与观测数据之间有更大的相关系数（CC）和更小的均方根误差（RMSE）。77.6%的站点数据集模拟与观测值之间的CC大于0.8，57.0%的台站大于0.9；80.4%的站点数据集模拟与观测值之间的RMSE小于3‰，56.1%的台站小于2‰。数据集的降水氧同位素时间序列的变化趋势与观测序列吻合较好，且明显优于iGCM的原始序列。同时，构建数据集的各月降水氧同位素空间格局与观测结果之间的一致性较好，展现了CNN融合模型的优势，即可以较好地利用各个模拟的优势，准确捕捉观测数据的特征。综上所述，构建的数据集能较好地再现观测数据，且具有时间连续和空间规则的特点，能为追踪大气和水文过程提供有效的数据支撑。然而，需要注意，在多数iGCM的共同时期（1969-2007），数据集可能更可靠。且受方法所限，在使用长序列数据时可能存在突变需要进行校正。此外，受观测站数据数量和代表性的影响，部分地区数据集的精度还有待提高。
This dataset contains stable oxygen isotopes in precipitation over mainland China from 1870 to 2017, with a spatial resolution of 50–60 km and a monthly temporal resolution. To fully utilize observational data and integrate the strengths of various introduced isotope-enabled Atmospheric General Circulation Models (iGCMs), a method combining data fusion and bias correction was adopted. Physically based auxiliary data including elevation and meteorological data (SRTM DEM, the 0.5° gridded monthly surface air temperature and precipitation data of China from the National Meteorological Information Center (NMIC)) were introduced into data fusion to enrich the climate and topographic information during the fusion process. The dataset was generated using different methods for each period: (1) 1870–1957: The simulation sequences from HadAM3 were corrected using two bias correction methods, linear scaling (LS) and quantile matching (DT), and the dataset adopted the ensemble mean of the two corrected results; (2) 1958–1968: The simulation sequences from CAM2 and HadAM3 were corrected using LS and DT, and the dataset adopted the ensemble mean of the two correction results for both sequences; (3) 1969–1978: The Convolutional Neural Network (CNN) fusion method was used to fuse four simulation sequences from three models (CAM2, GISS E free, GISS E nudged, and HadAM3) with the auxiliary data; (4) 1979–2001: The CNN fusion method was used to fuse nine simulation sequences from six models (CAM2, GISS E free, GISS E nudged, HadAM3, IsoGSM2, LMDZ4 free, LMDZ4 nudged, LMDZ4 zoomed, and MIROC32) with the auxiliary data; (5) 2002–2007: The CNN fusion method was used to fuse seven simulation sequences from four models (GISS E free, GISS E nudged, IsoGSM2, LMDZ4 free, LMDZ4 nudged, LMDZ4 zoomed, and MIROC32) with the auxiliary data; (6) 2008–2017: The simulation sequences from IsoGSM2 and LMDZ4 zoomed were corrected using LS and DT, and the dataset adopted the ensemble mean of the two correction results for both sequences. Prior to the dataset construction, the performance of three data fusion methods and two bias correction methods was compared. The results showed that the CNN-based fusion method performed the best, with a correlation coefficient (CC) with observational data greater than 0.95 and a root mean square error (RMSE) less than 1‰. The Back Propagation Neural Network (BPNN) and Long Short-Term Memory (LSTM) neural network exhibited similar performance, and both outperformed the bias correction methods. Therefore, the CNN fusion method was used to generate the dataset for the overlapping period of iGCMs from 1969 to 2007, while bias correction methods were applied for the remaining periods. Compared with iGCM simulations, the constructed dataset had a higher CC and smaller RMSE between the dataset and observational data. The CC between the simulated and observed values at 77.6% of the meteorological stations exceeded 0.8, and that at 57.0% of the stations exceeded 0.9; the RMSE between the simulated and observed values at 80.4% of the stations was less than 3‰, and that at 56.1% of the stations was less than 2‰. The temporal trend of the precipitation oxygen isotope time series in the dataset matched the observational sequence well, and was significantly better than the original iGCM sequences. Meanwhile, the spatial patterns of monthly precipitation oxygen isotopes in the constructed dataset were in good agreement with observational results, demonstrating the strengths of the CNN fusion model, which can effectively leverage the advantages of various simulations and accurately capture the characteristics of observational data. In summary, the constructed dataset can well reproduce observational data, and has the characteristics of temporal continuity and spatial regularity, providing effective data support for tracing atmospheric and hydrological processes. However, it should be noted that the dataset may be more reliable during the overlapping period of most iGCMs (1969–2007). Limited by the adopted method, abrupt changes may exist in the long time series and require post-correction. In addition, affected by the number and representativeness of observational stations, the accuracy of the dataset in some regions needs to be further improved.
