关键词: ERA5-Land, 再分析降水产品, 卷积神经网络, 分布式水文–土壤–植被模型, 径流模拟

Abstract:

Objective: This study performs convolutional neural network (CNN)-based residual correction of ERA5-Land precipitation data in subalpine forest regions of southwestern China and evaluates the applicability of the corrected data to hydrological modeling, aiming to provide a reference for precipitation data optimization and ecohydrological research in the region. Method: The upper reaches of Minjiang River in western Sichuan Province were selected as the study area, topographic factors and multiple meteorological variables were used as model inputs to construct a CNN-based residual correction model for ERA5-Land precipitation. Based on observations from 20 national meteorological stations in and surrounding the basin, a grid-to-station spatial matching method was used to evaluate the accuracy of the original and corrected ERA5-Land data and the IMERG, CMORPH-BLD, and CMA-RA/Land precipitation products. The distributed hydrological-soil-vegetation model (DHSVM) was then used to compare the runoff simulation effects driven by the different precipitation datasets and evaluate the applicability of the corrected data to hydrological modeling. Result: 1) Among the four original precipitation products, ERA5-Land exhibited relatively good overall performance, and its temporal variations were generally consistent with ground observations. However, it overestimated light precipitation, underestimated heavy precipitation, and overestimated total precipitation over the study area. 2) The CNN-based residual correction method substantially reduced the systematic bias of ERA5-Land while preserving its original data characteristics. For light precipitation, the percentage bias (PBIAS) of the corrected ERA5 Land data decreased from 104.94% to 33.92%. Across all precipitation intensity classes, the corrected ERA5-Land data exhibited higher correlation coefficient (CC) and lower normalized root mean square error (NRMSE) than the IMERG, CMORPH-BLD, and CMA-RA/Land products. 3) Among the DHSVM runoff simulations driven by the different precipitation datasets, the corrected ERA5-Land data produced the best performance, with an overall Nash–Sutcliffe efficiency (NSE) of 0.66 during the study period and relatively stable performance across years. 4) The IMERG, CMORPH-BLD, and CMA-RA/Land products underestimated winter snowfall over the study area to different extents, which delayed the simulated rise of spring snowmelt runoff. The corrected ERA5-Land data showed the smallest underestimation and the best spring snowmelt runoff simulation, with an NSE of 0.38, outperforming the other products. Conclusion: The CNN-based residual correction method effectively reduces the systematic bias in ERA5-Land precipitation and improves DHSVM runoff simulations driven by the corrected data. The corrected ERA5-Land data can provide a relatively reliable precipitation dataset for ecohydrological research in the sub-alpine forest regions of southwestern China.

Key words: ERA5-Land, reanalysis precipitation product, convolutional neural network, distributed hydrology soil vegetation model (DHSVM), runoff simulation