Due to its complex and diverse terrain, precipitation gauges in the Tibetan Plateau(TP) are sparse, making it difficult to obtain reliable precipitation data for environmental studies. Data merging is a method that ca...Due to its complex and diverse terrain, precipitation gauges in the Tibetan Plateau(TP) are sparse, making it difficult to obtain reliable precipitation data for environmental studies. Data merging is a method that can integrate precipitation data from multiple sources to generate high-precision precipitation data. However, the more commonly used methods, such as regression and machine learning, do not usually consider the local correlation of precipitation, so that the spatial pattern of precipitation cannot be reproduced, while deep learning methods do incorporate spatial correlation. To explore the ability of using deep learning methods in merging precipitation data for the TP, this study compared three methods: a deep learning method—a convolutional neural network(CNN) algorithm, a machine learning method—an artificial neural network(ANN) algorithm, and a statistical method based on Extended Triple Collocation(ETC) in merging precipitation from multiple sources(gauged, grid,satellite and dynamic downscaling) over the TP, as well as their performance for hydrological simulations. Dynamic downscaling data driven by global reanalysis data centered on the TP were introduced in the merging process to better reflect the spatial variability of precipitation. The results show that:(1) in terms of the meteorological metrics, the merged data perform better than the gauge interpolation data. By using data merging, the error between the raw multi-source and gauged precipitation can be reduced, and the precipitation detection capability can be greatly improved;(2) The merged precipitation data also perform well in the hydrological evaluation. The Xin’anjiang(XAJ) model parameter calibration experiments at the source of the Yangtze River(SYR) and the source of the Yellow River(SHR) were repeated 300 times to remove uncertainty in the model parameter results. The median Kling-Gupta Efficiency Coefficients(KGE) of simulated runoff from the merged data of the ANN, CNN and ETC methods for the SYR and the SHR are 0.859, 0.864, 0.838 and 0.835, 0.835, 0.789, respectively. Except for the ETC merging data at the SHR, the performance of other merged data was improved compared to the simulation results of the gauged precipitation(KGE=0.807 at the SYR, KGE=0.828 at the SHR);and(3) In contrast to the machine learning ANN method and the statistical ETC method, the deep learning method, CNN, consistently showed better performance.展开更多
基金supported by the National Natural Science Foundation of China(Grant No.52079093)the National Natural Science Foundation of Hubei Province of China(Grant No.2020CFA100)。
文摘Due to its complex and diverse terrain, precipitation gauges in the Tibetan Plateau(TP) are sparse, making it difficult to obtain reliable precipitation data for environmental studies. Data merging is a method that can integrate precipitation data from multiple sources to generate high-precision precipitation data. However, the more commonly used methods, such as regression and machine learning, do not usually consider the local correlation of precipitation, so that the spatial pattern of precipitation cannot be reproduced, while deep learning methods do incorporate spatial correlation. To explore the ability of using deep learning methods in merging precipitation data for the TP, this study compared three methods: a deep learning method—a convolutional neural network(CNN) algorithm, a machine learning method—an artificial neural network(ANN) algorithm, and a statistical method based on Extended Triple Collocation(ETC) in merging precipitation from multiple sources(gauged, grid,satellite and dynamic downscaling) over the TP, as well as their performance for hydrological simulations. Dynamic downscaling data driven by global reanalysis data centered on the TP were introduced in the merging process to better reflect the spatial variability of precipitation. The results show that:(1) in terms of the meteorological metrics, the merged data perform better than the gauge interpolation data. By using data merging, the error between the raw multi-source and gauged precipitation can be reduced, and the precipitation detection capability can be greatly improved;(2) The merged precipitation data also perform well in the hydrological evaluation. The Xin’anjiang(XAJ) model parameter calibration experiments at the source of the Yangtze River(SYR) and the source of the Yellow River(SHR) were repeated 300 times to remove uncertainty in the model parameter results. The median Kling-Gupta Efficiency Coefficients(KGE) of simulated runoff from the merged data of the ANN, CNN and ETC methods for the SYR and the SHR are 0.859, 0.864, 0.838 and 0.835, 0.835, 0.789, respectively. Except for the ETC merging data at the SHR, the performance of other merged data was improved compared to the simulation results of the gauged precipitation(KGE=0.807 at the SYR, KGE=0.828 at the SHR);and(3) In contrast to the machine learning ANN method and the statistical ETC method, the deep learning method, CNN, consistently showed better performance.