气温时序数据是全球和区域气候研究、水文模拟及生态系统评估中的关键基础数据。然而，尽管目前已存在多种气温数据产品，但由于观测站网分布不均、再分析数据与实测值之间的系统性偏差，以及长时间序列数据在空间覆盖一致性与高分辨率连续性方面的不足，当前对黄土高原长期气温动态与趋势变化的评估仍存在一定的不确定性。本研究利用ACCESS-ESM1-5、IPSL-CM6A-LR、MIROC-ES2L未来降水数据集，基于U-Net深度学习架构构建降尺度模型，生成黄土高原2015-2100年未来气温数据产品，空间覆盖范围为33.75°-41.25°N、100.95°-114.55°E，并具备0.1°的高空间分辨率与日时间分辨率。相较于原始模式数据，本数据集的核心优势在于其高分辨率气温数据产品更精细地刻画气温的时空分异规律，显著提升对局地极端高温、低温事件的模拟能力，从而为区域气候模拟、水文过程分析、及极端气温事件研究提供更可靠的数据支持。
\n本数据以NetCDF格式（*.nc）存储。数据文件命名格式为LoessPlateauRegion_ACCESS-ESM1-5_ssp126_temp_max_2015_2100_merged。其中，LoessPlateauRegion为地区名称，ACCESS-ESM1-5为模型名称，ssp126为未来情景名称。数据文件中temp_min为的日最低气温，temp_max为的日最高气温，时空分辨率分别为日、10 km，时间范围为2015-2100年，lat和long分别为模式格点的纬度和经度。", "ds_source": "
高分辨率观测数据采用基于欧洲中期天气预报中心第五代再分析数据（ERA5）开发的MSWX格点化偏差校正气象数据集，该数据集提供了10个近地表气象变量，空间分辨率为0.1°，时间分辨率为3小时，作为模型训练和验证的基准（https://www.gloh2o.org/mswx/）。基础气候数据来自耦合模式比较计划第六阶段（CMIP6）中3个全球气候模式的输出，包括历史时期（1950-2014年）和三种未来情景（SSP1-2.6, SSP2-4.5, SSP5-8.5，2015-2100年）的逐日模拟，其空间分辨率介于0.7°至2.8°之间（https://esgf-node.llnl.gov/search/cmip6/）
", "ds_process_way": "（1）利用MSWX数据集作为高分辨率观测数据，用于后续偏差校正与模型训练；（2）采用分位数Delta映射方法对CMIP6的3个全球气候模式输出进行偏差校正，校正后的数据用于降尺度输入；（3）基于U-Net深度学习架构构建降尺度模型，针对不同季节分别训练四个独立的U-Net模型，输入包括地形数据与偏差校正后的气象变量，并采用改进的损失函数以提升极端事件模拟能力；（4）利用1982–2014年期间的数据进行模型训练与验证，通过对比MSWX观测数据，从气候均值、极端事件频率及未来变化等多个方面系统评估降尺度数据集的性能。
", "ds_quality": "本研究采用均方根误差（RMSE）与平均偏差（MB）作为核心验证指标，系统评估了降尺度生成的日最高气温（tasmax）及日最低气温（tasmin）数据与MSWX高分辨率观测数据在1979年至2014年整个历史时期的一致性。
\n验证结果表明，降尺度后气温数据的精度得到根本性提升。具体而言，多模式集合平均的日最高气温（tasmax）RMSE从原始CMIP6输出的2.31°C以上普遍降至0.58°C以下，其MB范围也由-1.27至1.13°C改善至-0.52至-0.27°C；日最低气温（tasmin）的RMSE则由最优原始模式的2.32°C降至0.78°C。此外，数据能近乎完美地再现气温的季节循环，其多模式集合平均与观测的相关系数达1.0，且模型间不确定性显著降低。该降尺度气温数据在修正当前偏差的同时，也保留了原始模式预估的未来增温趋势，例如在SSP5-8.5情景下，其21世纪末的预估增温幅度（4.52°C）与原始模式（4.58°C）高度一致。因此，该降尺度气温数据集能够更可靠地反映区域气温的实际情况，可作为黄土高原高分辨率气候研究的可靠数据基础。
", "ds_acq_start_time": "2015-01-01 00:00:00", "ds_acq_end_time": "2100-12-31 00:00:00", "ds_acq_place": "黄土高原", "ds_acq_lon_east": 114.51666666666667, "ds_acq_lat_south": 33.75, "ds_acq_lon_west": 100.95, "ds_acq_lat_north": 41.25, "ds_acq_alt_low": null, "ds_acq_alt_high": null, "ds_share_type": "login-access", "ds_total_size": 23585065758, "ds_files_count": 0, "ds_format": "*.nc", "ds_space_res": "10km", "ds_time_res": "日", "ds_coordinate": "WGS84", "ds_projection": "", "ds_thumbnail": "9b6d4fec-a5b3-4ec7-8bf9-1cd85eeb5ab9.jpeg", "ds_thumb_from": 0, "ds_ref_way": "", "paper_ref_way": "", "ds_ref_instruction": "", "ds_from_station": "", "organization_id": "bf138922-7121-438c-8d1b-19d5f751c907", "ds_serv_man": "李红星", "ds_serv_phone": "0931-4967592", "ds_serv_mail": "ncdc@lzb.ac.cn", "doi_value": "", "subject_codes": [ "170.15" ], "quality_level": 0, "publish_time": "2026-05-21 09:55:41", "last_updated": "2026-05-21 10:49:52", "protected": false, "protected_to": null, "lang": "zh", "cstr": "11738.11.ncdc.loess.db7331.2026", "i18n": { "en": { "title": "Long-term daily composite temperature dataset for the Loess Plateau from 2015 to 2100", "ds_format": "*.nc", "ds_source": "The high-resolution observational data used in this study are based on the MSWX gridded bias-corrected meteorological dataset developed from the European Centre for Medium-Range Weather Forecasts (ECMWF) fifth-generation reanalysis (ERA5). This dataset provides 10 near-surface meteorological variables at a spatial resolution of 0.1° and a temporal resolution of 3 hours, serving as the reference dataset for model training and validation (https://www.gloh2o.org/mswx/). The baseline climate data were derived from outputs of three global climate models participating in the Coupled Model Intercomparison Project Phase 6 (CMIP6), including historical simulations (1950–2014) and daily projections under three future scenarios (SSP1-2.6, SSP2-4.5, and SSP5-8.5, 2015–2100). The spatial resolution of these climate models ranges from 0.7° to 2.8° (https://esgf-node.llnl.gov/search/cmip6/).", "ds_quality": "
This study employs Root Mean Square Error (RMSE) and Mean Bias (MB) as the core validation metrics to systematically evaluate the consistency between the downscaled daily maximum temperature (tasmax) and daily minimum temperature (tasmin) data and the high-resolution MSWX observational dataset over the entire historical period from 1979 to 2014.\r\nThe validation results demonstrate a fundamental improvement in the accuracy of the downscaled temperature data. Specifically, the multi-model ensemble mean RMSE for daily maximum temperature (tasmax) was generally reduced from above 2.31°C in the raw CMIP6 outputs to below 0.58°C, and its MB range improved from -1.27°C to 1.13°C to a narrower range of -0.52°C to -0.27°C. The RMSE for daily minimum temperature (tasmin) decreased from 2.32°C in the best-performing raw model to 0.78°C. Furthermore, the downscaled data can almost perfectly reproduce the seasonal cycle of temperature, with the multi-model ensemble mean showing a correlation coefficient of 1.0 against observations and a significant reduction in inter-model uncertainty. While correcting current biases, the downscaled temperature data also preserves the future warming trends projected by the original models. For example, under the SSP5-8.5 scenario, its projected warming magnitude by the end of the 21st century (4.52°C) is highly consistent with that of the raw models (4.58°C). Therefore, this downscaled temperature dataset can more reliably reflect actual regional temperature conditions and serves as a reliable data foundation for high-resolution climate research on the Loess Plateau.", "ds_ref_way": "", "ds_abstract": "
Temperature time series data serve as critical foundational information for global and regional climate research, hydrological modeling, and ecosystem assessments. However, despite the availability of various existing air temperature data products, the evaluation of long-term temperature dynamics and trend changes over the Loess Plateau remains subject to a certain degree of uncertainty. This is primarily due to challenges such as uneven distribution of observation networks, systematic biases between reanalysis data and measured values, and the lack of consistent spatial coverage and high-resolution continuity in long-term data series. To address these limitations, this study leverages future precipitation datasets from ACCESS-ESM1-5, IPSL-CM6A-LR, and MIROC-ES2L, constructing a downscaling model based on the U-Net deep learning architecture. The model is used to generate a future air temperature data product for the Loess Plateau covering the period 2015–2100. The spatial coverage of the dataset spans from 33.75°N to 41.25°N and 100.95°E to 114.55°E, with a high spatial resolution of 0.1° and daily temporal resolution. Compared to the original model data, the core advantage of this dataset lies in its ability to provide high-resolution temperature data that more finely captures the spatio-temporal distribution of air temperature. This significantly enhances the simulation of localized extreme high and low temperature events, thereby offering more reliable data support for regional climate modeling, hydrological process analysis, and studies of extreme temperature events.\r\n
The data is stored in NetCDF format (*.nc). The data files are named using the format: LoessPlateauRegion_ACCESS-ESM1-5_ssp126_temp_max_2015_2100_merged. In this naming convention, LoessPlateauRegion denotes the regional name, ACCESS-ESM1-5 specifies the climate model name, and ssp126 indicates the future scenario. Within the data files, the variable temp_min represents daily minimum temperature, and temp_max represents daily maximum temperature. The spatial and temporal resolutions are 10 km and daily, respectively, covering the period from 2015 to 2100. The lat and lon fields correspond to the latitude and longitude of the model grid points.", "ds_time_res": "", "ds_acq_place": "Loess Plateau", "ds_space_res": "", "ds_projection": "", "ds_process_way": "
(1) The MSWX dataset is utilized as high-resolution observational data for subsequent bias correction and model training.(2) The Quantile Delta Mapping method is applied to perform bias correction on the outputs of three CMIP6 global climate models. The corrected data are then used as inputs for downscaling.(3) A downscaling model is constructed based on the U-Net deep learning architecture. Four independent U-Net models are trained separately for different seasons, incorporating topographic data and bias-corrected meteorological variables as inputs. An improved loss function is employed to enhance the simulation capability for extreme events.(4) Data from the period 1982–2014 are used for model training and validation. The performance of the downscaled dataset is systematically evaluated by comparing it with MSWX observational data across multiple aspects, including climatic means, frequency of extreme events, and future changes.", "ds_ref_instruction": "" } }, "submit_center_id": "ncdc", "data_level": 0, "recommendation_value": 0, "license_type": "https://creativecommons.org/licenses/by/4.0/", "doi_reg_from": "reg_local", "cstr_reg_from": "reg_local", "doi_not_reg_reason": null, "cstr_not_reg_reason": null, "is_paper_in_submitting": false, "belong_to_nieer": false, "ds_topic_tags": [ "气温", "长时序", "深度学习", "高分辨率" ], "ds_subject_tags": [ "大气科学" ], "ds_class_tags": [], "ds_locus_tags": [ "中国", "黄土高原" ], "ds_time_tags": [ 2015, 2016, 2017, 2018, 2019, 2020, 2021, 2022, 2023, 2024, 2025, 2026, 2027, 2028, 2029, 2030, 2031, 2032, 2033, 2034, 2035, 2036, 2037, 2038, 2039, 2040, 2041, 2042, 2043, 2044, 2045, 2046, 2047, 2048, 2049, 2050, 2051, 2052, 2053, 2054, 2055, 2056, 2057, 2058, 2059, 2060, 2061, 2062, 2063, 2064, 2065, 2066, 2067, 2068, 2069, 2070, 2071, 2072, 2073, 2074, 2075, 2076, 2077, 2078, 2079, 2080, 2081, 2082, 2083, 2084, 2085, 2086, 2087, 2088, 2089, 2090, 2091, 2092, 2093, 2094, 2095, 2096, 2097, 2098, 2099, 2100 ], "ds_contributors": [ { "true_name": "张宝庆", "email": "baoqzhang@lzu.edu.cn", "work_for": "兰州大学", "country": "中国" } ], "ds_meta_authors": [ { "true_name": "张宝庆", "email": "baoqzhang@lzu.edu.cn", "work_for": "兰州大学", "country": "中国" } ], "ds_managers": [ { "true_name": "张宝庆", "email": "baoqzhang@lzu.edu.cn", "work_for": "兰州大学", "country": "中国" } ], "category": "气象" }