1.本数据集面向青藏高原长江、黄河源区复杂地形区域的高精度气象驱动需求，提供长江、黄河源区区域尺度上的逐日2米近地气温的nc格式数据，空间分辨率为4.0 km（约 0.03333°），时间范围覆盖1960-01-01至2023-12-31。研究地点位于青藏高原腹地，被称为“中华水塔”，主要分布在青海省的唐古拉山和巴颜喀拉山脉一带。该区域平均海拔约4500m，气候寒冷干燥，冰川和多年冻土广泛分布。黄河源区位于黄河干流龙阳峡水库上方，集中在巴颜喀拉山北麓的约古宗列盆地及扎陵湖、鄂陵湖附近，地理坐标介于 95°55'E-98°41'E、33°56'N-35°31'N 之间。属典型的大陆性高原气候，年平均气温大致在-3 ℃~ -4.1 ℃，年降雨量通常在300~700 mm。长江源区以直门达水文站为界，位于唐古拉山和昆仑山之间，范围在 90°43′E-97°45′E，32°30′N-36°35′N之间，整体气候干冷少雨，年均气温-1.7 ℃~-5.5 ℃，年均降水量约为270~410 mm。\n
研制流程如下：首先，开展了为期2年(1960和1961年)空间分辨率为1/30°的WRF模拟；其次，利用WRF模拟结果在日尺度上分别训练基于卷积神经网络CNN的降尺度模型。该降尺度模型由四个卷积层(用于特征提取)和一个亚像元卷积层(Subpixel convolution，用于构建高分辨率数据)组成。模型的输入包括粗分辨率气温数据、粗分辨率地形数据(即网格内的海拔和海拔标准差)及高分辨率地形数据，输出为高分辨率气象数据；然后，利用训练好的模型对长时间序列ERA5再分析数据进行降尺度，以生成高分辨率(1/30°≈4 km)的格点气温数据（ERA5_CNN），并利用源区台站观测实施CDF（累积分布函数）偏差订正，最后裁剪到长江—黄河源区边界生成区域产品。数据严格遵循 CF-1.8 / ACDD-1.3 元数据规范，提供 NetCDF-4格式文件，既便于python科学计算，也能在 ArcGIS Pro 中直接加载。\n
2.数据内容与要素： 变量：T_2m——2 m 逐日平均气温（°C）。\n
3.时空范围：长江黄河源区（经度 90.55–103.41°E，纬度 32.15–36.11°N）；时间 1960-01-01 至 2023-12-31。\n
4.分辨率：空间 0.03333° × 0.03333°（约 4 km），时间 日。\n
5.命名方法：YYYY_MM_DD.nc。\n
6.坐标：latitude、longitude；CRS：EPSG:4326。\n
7.生产背景与方法概述：青藏高原地形起伏大、下垫面差异显著，常规 0.25° 再分析或简单插值方法难以刻画局地热力—地形效应。本数据集利用 WRF 短期高分辨率模拟学习“高分辨率—低分辨率”的映射关系，并以 CNN 捕捉非线性空间特征，在全时段对 ERA5 进行统计降尺度；随后以源区台站观测实施 CDF订正，显著削弱高海拔冷偏差，提升站点一致性。\n
8.优势与特点： ①更高空间分辨率与地形细节保持（≈4 km vs 0.25°）； ②物理先验（动力降尺度） + 深度学习（统计降尺度）的组合，优于仅插值的空间锐化； ③偏差订正后与台站更一致，极值表现与年际变化更可信；\n
9.应用范围：多年冻土/活动层热状况评估、冻融指数与 N 因子计算、流域水文与生态模型驱动、区域气候变化检测、地表过程模拟与灾害风险评估等。", "ds_source": "
1.再分析数据（背景场） ERA5 2 m 气温（ECMWF/Copernicus），时空覆盖长、同化方案先进，提供大尺度逐日/逐小时近地层气象场。原始分辨率约 0.25°，小时尺度；在本工作中先聚合为日均作为 CNN 的输入特征之一。\n
\n2.短期高分辨率大气模拟（高分辨率特征） WRF（ARW） 短期模拟覆盖 1960 年全年 和 1961 年 6–9 月关键季节，提供近地气温及地形相关高分辨率结构信息,用于构建目标尺度（公里级）温度“训练”样本，；用于训练 CNN 学习 Low-High 分辨率映射关系与复杂地形下的空间特征。
\n3.台站观测（订正与验证）用于独立验证与不确定性评估（来自中国气象局（CMA）的逐日常规气象要素观测）,（源区内及周边共 28 个站，其中 15 个站 1960 年起始），用于：a) 训练/验证集划分与交叉检验；b) CDF偏差订正的统计映射构建；c) 独立评估（RMSE/偏差/相关系数等）4.其他辅助数据 4 km DEM （地形标准差等，可作为 CNN 的静态特征输入模型）。
", "ds_process_way": "1.预处理：对背景场0.25°分辨率的ERA5数据进行裁剪，裁剪出包含长江、黄河源区的矩形范围；统一到 WGS84 网格；异常值剔除与基本质控。\n
2.动力—统计降尺度：使用 WRF在代表期进行高分辨率模拟，结合地形等静态因子，训练 CNN模型 将 ERA5低分辨率温度映射到公里级；\n
3.偏差订正：使用CDF分位数映射方法对降尺度结果进行偏差校正；\n
4.数据写出标准与元数据生成方法：按 CF-1.8/ACDD-1.3 标准生成 NetCDF4 文件。本数据以“再分析基础场 + 短期高分辨率模拟 + CNN 统计降尺度 + CDF 偏差订正”的流程生产： (1) 预处理：将 ERA5 统一为日均；对坐标、时区进行一致化； (2) 短期高分辨率模拟：以 WRF 短期高分辨率模拟（1960 全年 + 1961 年关键季节6-9月）获取复杂地形下的高分辨率场，用作 CNN 训练的目标特征； (3) CNN 降尺度：构建以 ERA5 及静态地形/WRF 短期低分辨率模拟为输入、以高分辨率气温为目标的 CNN 模型；训练后对 1960–2023 全时段进行逐日降尺度，得到 0.03333° 分辨率气温场； (4) CDF 偏差订正：以台站观测构建“观测—降尺度”逐站 CDF 映射，对网格进行订正（站点→反距离加权空间插值）； (5) 质控与裁剪：数值合理性（[-50, 50] °C）、时间连续性、CF/ACDD/CRS 元数据完善，裁剪至源区边界并输出 NetCDF-4。 （参考文献： 1.Jiang, Y., et al., A downscaling approach for constructing high-resolution precipitation dataset over the Tibetan Plateau from ERA5 reanalysis. Atmospheric Research, 2021. 256: p. 105574. 2.姜尧志等, 第三极地区高分辨率近地面气象驱动数据研制. 中国科学:地球科学, 2025. 55(04): 第1320-1337页. 3.Zhou, X., et al., Added value of kilometer-scale modeling over the third pole region: a CORDEX-CPTP pilot study. Climate Dynamics, 2021. 57(7-8): p. 1-15. 4.Zhou, P., et al., High resolution Tibetan Plateau regional reanalysis 1961-present. Scientific Data, 2024. 11(1): p. 444. ）\n3.Zhou, X., et al., Added value of kilometer-scale modeling over the third pole region: a CORDEX-CPTP pilot study. Climate Dynamics, 2021. 57(7-8): p. 1-15.\n4.Zhou, P., et al., High resolution Tibetan Plateau regional reanalysis 1961-present. Scientific Data, 2024. 11(1): p. 444.\n）\n", "ds_quality": "
1.结构与规范：文件包含完备的坐标、单位与全局元数据；
\n2.验证与不确定性：①将用于研制数据集的背景场数据：0.25° ERA5日均温数据采用最常用的双线性插值方法（bilinear interpolation）将0.25°的ERA5日均温插值到4km分辨率，用作方法对比的基线；②然后利用28个独立台站在区域内对两个数据集进行对比验证（RMSE/MAE/Bias/相关性等指标），深度学习降尺度产品 ERA5_CNN 相比双线性插值 ERA5_BLI 显著降低了系统性冷偏差与均方误差（区域平均 Bias 从 -2.72 ℃ 降至 0.01 ℃，RMSE降低了超过50%，从 4.41 ℃ 降至 1.97 ℃），且每个站点的相关系数（CC）普遍保持在0.9以上。
\n综上，CNN 降尺度方法在源区复杂地形内的逐日气温再现能力显著优于简单的双线性插值，更接近观测值。
\n一些局限性：在极端气候寒冷或高风条件下可能存在轻微系统性偏差。不确定性来源包括再分析本身偏差、CNN 对极端值的回归到均值、台站稀疏导致的局地代表性等。\n不确定性来源包括再分析本身偏差、CNN 对极端值的回归到均值、台站稀疏导致的局地代表性等。
", "ds_acq_start_time": "1960-01-01 00:00:00", "ds_acq_end_time": "2023-12-31 00:00:00", "ds_acq_place": "长江与黄河源头地区经纬度范围约为32°N-36°N、89°E-103°E", "ds_acq_lon_east": 103.41305555555556, "ds_acq_lat_south": 36.114444444444445, "ds_acq_lon_west": 90.54777777777778, "ds_acq_lat_north": 32.14805555555556, "ds_acq_alt_low": 2662.0, "ds_acq_alt_high": 6479.0, "ds_share_type": "open-access", "ds_total_size": 2104026559, "ds_files_count": 23380, "ds_format": "NetCDF", "ds_space_res": "4千米", "ds_time_res": "日", "ds_coordinate": "WGS84", "ds_projection": "无", "ds_thumbnail": "b96920e8-bf9d-48ae-ab84-f77fd0afea8d.png", "ds_thumb_from": 0, "ds_ref_way": "", "paper_ref_way": "", "ds_ref_instruction": "可以通过ARCGIS和python进行可视化", "ds_from_station": null, "organization_id": "52b7b79b-860c-49a5-9083-9a70cf8bed5a", "ds_serv_man": "李红星", "ds_serv_phone": "0931-4967592", "ds_serv_mail": "ncdc@lzb.ac.cn", "doi_value": "", "subject_codes": [ "170.15", "170.1535", "170.1525" ], "quality_level": 3, "publish_time": "2025-11-07 15:19:57", "last_updated": "2026-04-09 16:54:08", "protected": false, "protected_to": null, "lang": "zh", "cstr": "11738.11.NCDC.NIEER.DB7004.2025", "i18n": { "en": { "title": "Daily 2-meter air temperature dataset for the Yangtze and Yellow River source regions at 4-kilometer resolution, 1960–1979", "ds_format": "", "ds_source": "1. Reanalysis Data (Background Field) ERA5 2 m Temperature (ECMWF/Copernicus) features extensive spatiotemporal coverage and advanced assimilation schemes, providing large-scale daily/hourly near-surface meteorological fields. With native resolution of approximately 0.25° at hourly intervals, this dataset is aggregated to daily averages in this work as one of the CNN input features.\n
\n2. Short-term high-resolution atmospheric simulation (high-resolution features) WRF (ARW) short-term simulations cover the full year of 1960 and the critical season of June–September 1961, providing high-resolution structural information on near-surface temperature and topography. This data is used to construct target-scale (kilometer-level) temperature “training” samples for constructing CNNs to learn low-high resolution mapping relationships and spatial features under complex terrain.\n
\n3. Station Observations (Correction and Validation) Used for independent validation and uncertainty assessment (daily conventional meteorological observations from the China Meteorological Administration (CMA)), sourced from 29 stations within and adjacent to the source region (15 stations with data starting from 1960). Applications include: a) Training/validation set partitioning and cross-validation; b) Statistical mapping construction for CDF bias correction; c) Independent evaluation (RMSE/bias/correlation coefficients, etc.)\n
\n4. Other Auxiliary Data 4 km DEM (terrain standard deviation, etc., usable as static feature input for CNN models);", "ds_quality": "
1. Structure and Specifications: The file contains complete coordinates, units, and global metadata. \n
\n2. Validation and Uncertainty: ① The background field data used to develop the dataset—0.25° ERA5 daily mean temperature data—was interpolated to 4 km resolution using the most common bilinear interpolation method. This serves as the baseline for method comparison. ② Subsequently, two datasets were validated against each other using 29 independent stations within the region (using metrics such as RMSE/MAE/Bias/correlation). The deep learning downscaling product ERA5_CNN significantly reduced systematic cold bias and mean squared error compared to bilinear interpolation ERA5_BLI (regional average Bias decreased from -2.74 ℃ to -0.01 ℃, RMSE decreased by over 50%, from 4.48°C to 2.04°C), while the correlation coefficient (CC) at each station generally remained above 0.9.\n
\nIn summary, the CNN downscaling method significantly outperforms simple bilinear interpolation in reproducing daily temperatures within complex terrain in source regions, yielding results closer to observations. Some limitations exist: minor systematic biases may occur under extreme cold or high wind conditions. Sources of uncertainty include biases inherent in the reanalysis itself, CNN regression toward the mean for extreme values, and local representativeness issues due to sparse station coverage.
", "ds_ref_way": "", "ds_abstract": "1. This dataset addresses the need for high-precision meteorological drivers in the complex terrain regions of the Yangtze and Yellow River headwaters on the Qinghai-Tibet Plateau. It provides daily 2-meter near-surface air temperature data in NC format at the regional scale for the Yangtze and Yellow River headwaters, with a spatial resolution of 4.0 km (approximately 0.03333°). The temporal range covers January 1, 1960, to December 31, 1979.The study site is located in the heart of the Qinghai-Tibet Plateau, known as the “Water Tower of China,” primarily distributed in the Tangula and Bayan Har mountain ranges of Qinghai Province. This region has an average elevation of approximately 4,500 meters, featuring a cold, arid climate with widespread glaciers and permafrost. The Yellow River headwaters region lies upstream of the Longyangxia Reservoir on the main stem of the Yellow River. It is concentrated in the Yoguzonglie Basin on the northern foothills of the Bayankala Mountains and near Lake Zhaling and Lake Eling, with geographical coordinates between 95°55'E-98°41'E and 33°56'N-35°31'N. It exhibits a typical continental plateau climate, with an average annual temperature ranging from approximately -3°C to -4.1°C and annual precipitation typically between 300–700 mm. The Yangtze River source region, demarcated by the Zhimen Da hydrological station, lies between the Tanggula and Kunlun Mountains. Its boundaries span 90°43′E-97°45′E and 32°30′N-36°35′N. The overall climate is dry, cold, and low in precipitation, with an average annual temperature of -1.7°C to -5.5°C and annual precipitation of approximately 270–410 mm.\n
The development process is as follows: First, a two-year (1960 and 1961) WRF simulation with a spatial resolution of 1/30° was conducted. Second, a downscaling model based on a convolutional neural network (CNN) was trained at the daily scale using the WRF simulation results. This downscaling model comprises four convolutional layers (for feature extraction) and one subpixel convolution layer (for constructing high-resolution data). Model inputs include coarse-resolution air temperature data, coarse-resolution terrain data (i.e., grid-based elevation and elevation standard deviation), and high-resolution terrain data. The output is high-resolution meteorological data. The trained model is then applied to downscale long-term ERA5 reanalysis data, generating high-resolution (1/30°≈4 km) gridded air temperature data (ERA5_CNN). Cumulative distribution function (CDF) bias correction is implemented using source region station observations, and the final regional product is cropped to the boundaries of the Yangtze-Yellow River source area. Data strictly adheres to CF-1.8 / ACDD-1.3 metadata specifications, provided as NetCDF-4 format files. This format facilitates Python scientific computing and enables direct loading in ArcGIS Pro.\n
\n2. Data Content and Elements: Variable: T_2m(time, lat, lon) — 2 m daily mean air temperature (°C), standard_name=air_temperature; associated scalar coordinate height=2.0 m; grid_mapping=crs(WGS84).\n
\n3. Spatial-Temporal Extent: Longitude 90.547953–103.413333°E, Latitude 32.148290–36.114560°N; Time 1960-01-01 to 1979-12-31 (including leap years).\n
\n4. Resolution: Spatial 0.03333° × 0.03333° (approx. 4 km), Temporal Daily.\n
\n5. Naming convention: Daily file ERA5_CNN_t2m_4km_daily_YYYYMMDD.nc (time=1).\n
\n6. Coordinates: lat/degN, lon/degE in ascending order; CRS: EPSG:4326.\n
\n7. Production Background and Method Overview: The Qinghai-Tibet Plateau exhibits significant topographic variations and distinct surface conditions, making it challenging for conventional 0.25° reanalysis or simple interpolation methods to capture local thermodynamic-topographic effects. This dataset employs WRF short-term high-resolution simulations to learn the “high-resolution to low-resolution” mapping relationship, using CNN to capture nonlinear spatial features for statistical downscaling of ERA5 across all time periods. Subsequently, CDF corrections are applied using source region station observations (site-scale mapping constructed and IDW spatially interpolated), significantly reducing high-altitude cold biases and enhancing station consistency.\n
\n8. Advantages and Features: ① Higher spatial resolution with terrain detail preservation (≈4 km vs 0.25°); ② Combination of physical prior (dynamic downscaling) + deep learning (statistical downscaling) outperforms interpolation-only spatial sharpening; ③ Improved station consistency after bias correction, yielding more reliable extreme events and interannual variability;\n
\n9. Applications: Permafrost/active layer thermal status assessment, frost-thaw index and N-factor calculation, watershed hydrological and ecological modeling, regional climate change detection, surface process simulation, and disaster risk assessment.Translated with DeepL.com (free version)
", "ds_time_res": "日", "ds_acq_place": "The latitudinal and longitudinal ranges of the Yangtze and Yellow River headwaters are approximately 32°N–36°N and 89°E–103°E.", "ds_space_res": "4千米", "ds_projection": "", "ds_process_way": "Preprocessing: Clipping ERA5 data with 0.25° resolution for the background field to extract a rectangular area encompassing the Yangtze and Yellow River source regions; unifying to WGS84 grid; outlier removal and basic quality control.