Considering that large lakes are likely to block the transport of materials from upstream to downstream, two types of catchments – full catchments and inter-lake catchments – are defined in this study. The full catchment refers to the full upstream-contributing area of each lake, while the inter-lake catchment is obtained by excluding the contributing areas of upstream lakes larger than 0.2 km2 from the full catchment following the definition by Soranno et al. (2017). For example, the green area in Fig. 3 is the full watershed of lake No. 1, and the stippled black area is its inter-lake catchment.

Traditional river-oriented catchment delineation methods are not suitable for the delineation of lake catchments. Liu et al. (2020) proposed a lake-oriented approach to delineating endorheic catchments, which can be used to delineate the full catchments of endorheic lakes in this study. But there are more tasks in this study, including the delineation of both full catchments and inter-lake catchments for endorheic lakes and upstream lakes, the construction of topological relationships among lakes/lake-catchments, and the tracing of flow paths among upstream and stream lakes. Therefore, we developed a software using the C and Python programming language to implement these tasks; the source code is open (https://github.com/LoserOne-ovo/basin_delineation, last access: 18 August 2022).

Flow direction and lake boundary data are needed for catchment delineation. Firstly, the vector lake data are rasterized using the same geospatial reference system and pixel size as the flow direction data (Yamazaki et al., 2019). Concurrently, the reverse-flow-direction data are calculated (recorded as 8 bits corresponding to 8 neighbors, e.g., 10000001 means the first and eighth neighbors flow into the current pixel) to assist the tracing of upstream contributing areas. Next, this is iterated over all the pixels to find the inlets to the lake, which are defined as the pixels flowing directly into a lake according to the flow direction. Then the inlet pixels are put into a stack. This is repeated by taking a pixel out and pushing its direct upstream pixels in until the stack is empty. Finally, we get the full upstream contributing area (i.e., full catchment) of the lake. Following the same procedures but adding a termination criterion that if a tracing branch meets an upstream lake pixel then return, the inter-lake catchments can be obtained.

In this study, flow direction data from the MERIT Hydro dataset (Yamazaki et al., 2019) were used, which have a resolution of 3 arcsec (∼ 90 m at the Equator). This dataset was verified to have high accuracy in terms of flow accumulation area and river basin shape. The lake boundaries in 2018 were used in the delineation of catchments, and the dataset from Zhang et al. (2019) (https://doi.org/10.11888/Hydro.tpdc.270303, Zhang, 2019​​​​​​​) was adopted. Some additional operations such as merging adjacent lakes and filling holes were employed to remove topology errors.

The lake catchments are characterized through six aspects: lake body, topography, climate, land cover/use, soil and geology, and anthropogenic activities. The following standards were adopted for source dataset selection. Firstly, the corresponding environmental variables should have theoretical impacts on the attributes of lake water or sediment. Secondly, the dataset covers the whole TP and can be publicly available. Finally, the most reputable datasets with high resolution are preferred when multiple data sources are available. Following these standards, the vast majority of selected datasets were well validated and recognized in each field. For example, the SoilGrids dataset (Poggio et al., 2021) is the de facto standard of gridded soil property data, and the China meteorological forcing dataset (CMFD) (Yang and He, 2019) is the most widely used climatic dataset on the TP. The information of source datasets used in the LCC-TP dataset can be found in Table S1 in the Supplement. All the spatial datasets were converted to GDAL/OGR readable formats and projected to the Albers equal-area conic projection. Besides static attributes, the dynamic time series of hydrological and meteorological data are also provided. Version 1.0 of LCC-TP offers 57 different variables with a total of 721 individual attributes. The details of these attributes are described in Sect. 4.

For each lake, zonal statistical analyses are conducted to calculate catchment-level characteristics for inter-lake catchments and full catchments. Different methods are used for raster data and vector data. For raster data, the grid cells within each catchment are picked out and the statistics of their values are calculated. For continuous variables (e.g., elevation and precipitation amount), the average value or maximum/minimum value is calculated. For categorical variables (e.g., land use type), the mode is calculated, and the percentages of each category are also calculated for some variables. The rasterstats Python package (https://github.com/perrygeo/python-rasterstats, last access: 18 August 2022) and the Google Earth Engine cloud platform were adopted for implementation. Vector data, such as the spatial distribution of glaciers stored in the ESRI shapefile format, are first intersected with the catchment extent layer, then the ratio of the intersection's area within each catchment to the catchment's total area is calculated. The spatial distribution of the catchment-scale variables obtained was plotted and visually checked by the authors to ensure the correctness of the dataset. For numerical variables whose original data were in raster format, their ranges at the grid-cell scale (i.e., before zonal statistics) and the catchment scale (i.e., after zonal statistics) were calculated (Table S2), and we checked that the range at the catchment scale fell within that at the grid-cell scale for each variable.

The catchment delineation in this study was based on the flow-direction data from Yamazaki et al. (2019), which have been widely verified. To further validate the accuracy of the delineated catchments, the dataset from Liu et al. (2020) was used as a reference, which contains the boundaries of 421 lake catchments on the Inner TP. Figure 5 shows that the catchment areas in this study have a high correlation (r=0.988) with those of Liu et al. (2020), which proves the correctness of our results. The small differences between these two datasets may be related to the errors in the DEM and the different methods for depression filling and flow direction correction.

The area, perimeter, lake development index, and type of lake were calculated. The lake development index was used to characterize the complexity of the lake shoreline, which was defined in formula (1), where L represents the length of the lake shoreline and s represents lake area, i.e., the ratio of the shoreline length to the circumference of a circle with the same area as the lake. The value increases with increasing shoreline complexity, and the maximum value is 1 while the shape is a circle: 1dev=L2πs. The type of lake herein refers to whether it is an upstream lake (i.e., a lake with outflow to a downstream river or lake) or a terminal lake (i.e., a lake without outflows).

The smallest lake has an area of 0.2 km2 and the largest lake (i.e., Qinghai Lake) has an area of 4503.5 km2. Overall, 72 % of the 1525 lakes have an area less than 10 km2. Figure 6 shows the spatial distribution of the lake development index and type on the TP. The average development index for lakes across the TP is 3.40, and there is a cluster of lakes with high development indices in the north of the Inner TP. Out of 1525 lakes, 364 (24 %) are terminal lakes, most of which are located in the Inner TP and the Qaidam Basin.

The catchment-level elevation (including average, maximum, and minimum value), relief, slope, catchment area, and lake-catchment area ratio were calculated based on the MERIT DEM (Yamazaki et al., 2017). These characteristics were calculated for both the full and inter-lake catchments. The slope values were calculated using ArcGIS 10.5. The relief values, defined as the difference between the maximum and minimum elevations in a neighborhood, were calculated using window sizes of 5×5, 11×11, 21×21, 31×31, 41×41, and 51×51 based on a DEM of 0.00833∘ resolution. The usage of different window sizes in the calculation of the relief value aims to meet the needs of different analysis scenarios: for the research focusing on small-scale terrain variation, a small window size is appropriate; when the focus is large-scale terrain variation, a larger window size is preferred. Figure 7 shows the spatial distribution of mean elevation, relief, slope (%), and lake-catchment area ratio for inter-lake catchments on the TP. The mean elevation is relatively low in the eastern TP and the valley between the Kunlun and Gangdise mountains in the south of the Inner TP. The relief and slope are relatively low in the north of the Inner TP, where the elevation is very high. The lake-catchment area ratio is high in the south and east parts of the Inner TP and the upper Yellow River basin.

A total of 11 climatic variables were included in the constructed dataset, including 2 m air temperature, surface pressure, and specific humidity, 10 m wind speed, downward shortwave radiation, downward longwave radiation, precipitation amount, potential evapotranspiration (PET), actual evapotranspiration (AET), climate moisture index (CMI), and aridity index. The multi-year average values of all the variables were calculated at three levels (i.e., the lake body, inter-lake catchment, and full catchment level), and the monthly and growing-season (May–September) average values of all the variables except the aridity index were also calculated.

The grid-based CMFD dataset (Yang and He, 2019), ranging from 1979 to 2018, was used to calculate the catchment-level climatic characteristics. CMFD was constructed through the fusion of in situ observations from weather stations, remote-sensing products, and reanalysis datasets, which improved the data quality in western China where weather stations are sparse. It has a spatial resolution of 0.1∘ and a temporal resolution of 3 h. The background field data of air temperature came from GLDAS NOAH10SUBP 3H V001 and the precipitation data were the combination of GLDAS NOAH10SUBP 3H V001, GLDAS NOAH025 3H V2.1, and TRMM 3B42 V7.

PET was derived from the Global Potential Evapotranspiration (Global-PET) dataset (Zomer et al., 2008). In this dataset, monthly PET was estimated via the Hargreaves (1994) equation at a spatial resolution of 30 arcsec using precipitation and temperature inputs obtained from the WorldClim dataset (Hijmans et al., 2005). The aridity index was derived from the Global-Aridity dataset (Zomer et al., 2008), which quantifies precipitation availability over atmospheric water demand and was calculated as the ratio of long-term mean precipitation and PET. CMI was another metric to characterize the degree of humidity, which is defined via the following function: [CMI=(P/PET)-1 when P<PET] or [CMI=1-(PET/P) when P≥PET​​​​​​​] (Willmott and Feddema, 1992).

Figure 8 shows the spatial distribution of multi-year average climatic characteristics for inter-lake catchments on the TP. The air temperature and pressure are low in the north of the Inner TP where elevation is high. Radiation is high in the southwest of the TP and low in the north part. Wind speed is high in the east and southwest of the Inner TP. Precipitation and evapotranspiration have a decreasing trend from southeast to northwest, and accordingly it gets drier from southeast to the northwest as shown by the spatial distribution of air specific humidity, climate moisture index, and aridity index. It should be noted that aridity indices are higher under more humid conditions and lower under more arid conditions according to its formula.

The land cover/use characteristics include remote-sensed vegetation indices (i.e., EVI (enhanced vegetation index) and NDVI (normalized difference vegetation index)), gross primary productivity (GPP), net primary production (NPP), and dominant land cover/use type in each catchment as well as the fractions of each type, and the fractions of protected area. The average vegetation indices, GPP, and NPP across the whole year and in the growing season were calculated.

The land cover/use data came from the fusion land use product on the TP produced by Xu (2019), which was constructed based on six mainstream land use products, i.e., ESA GlobCover (Arino and Bicheron, 2010​​​​​​​), NLCD-China (Liu et al., 2005), FROM-GLC (Gong et al., 2019), MODIS MCD12Q1 (Friedl and Sulla-Menashe, 2019), UMD GLCF GLCDS (Hansen et al., 1998), and USGS IGBP DISCover (Loveland et al., 2009). This dataset has a spatial resolution of 300 m and covers three historical periods (1992, 2005, and 2015). There are nine different land use types over the TP, including grassland, shrubland, forest, glacier, bare land, water body and wetland, desert, farmland, and urban land. In addition, the second glacier inventory dataset of China (version 1.0, 2006–2011) (Liu et al., 2012) and a wetland distribution dataset (the 1970s, 2000s) (Zhou, 2018) were used as independent datasets for glaciers and wetlands. For the protected area, the World Database on Protected Areas (WDPA) (UNEP-WCMC and IUCN, 2021) was used. The fractional snow cover data (i.e., the fraction of a pixel that is snow covered) was extracted from the MODIS daily cloud-free snow cover product over the TP (2002–2015) (Qiu, 2018a). Considering that cloud and snow have similar reflection signals, eight different methods were employed in this product to remove the influence of cloud on snow cover identification.

Figure 9 shows the spatial distribution of land use/cover characteristics for inter-lake catchments on the TP. EVI is generally low for catchments across the TP and there is a deceasing trend from southeast to northwest following the spatial pattern of precipitation. The fraction of grassland is higher in the south of the Inner TP and the source region of the Yellow River, and that of shrubland is higher in the middle of the Inner TP. Wetlands have higher coverage mainly in the south and east part of the Inner TP and the upper Yellow River basin. There is no cropland in most lake catchments, and the bare land and desert are mainly distributed in the north of the TP. The fraction of glaciers is relatively higher in the south and west of the TP.

This study included 21 physical and chemical variables of soil. The proportions of sand, silt, clay, and coarse fragments, the bulk density, cation exchange capacity (CEC), pH, total nitrogen (TN), and soil organic carbon (SOC) content/density/stock were derived from the 250 m resolution SoilGrids 2.0 product (Poggio et al., 2021). These soil properties were predicted at six different depths (i.e., 0–5, 5–15, 15–30, 30–60, 60–100, and 100–200 cm) using machine learning models based on observations from over 230 000 soil profiles globally in the WoSIS database and over 400 environmental covariates. Data of all the six layers were aggregated at the inter-lake catchment and full catchment level. The grading standards of soil particle size are as follows: coarse fragments in the soil refer to particles larger than 2 mm and smaller than 25 cm in diameter, sand refers to particles >0.05 mm in the fine earth fraction (i.e., the particles less than 2 mm), silt refers to particles ≥0.002 and ≤0.05 mm in the fine earth fraction, and clay refers to the particles <0.002 mm. The CEC measures the ability of soil to hold onto exchangeable cations which can resist nutrient leaching and it is an important variable in soil fertility. The organic carbon content, density, and stock are the mass of organic carbon per unit mass, volume, and surface area, respectively. Besides SoilGrids, the dataset of SOC content at different depth intervals (i.e., 0–30, 0–50, 0–100, 0–200, and 0–300 cm) over the Third Pole produced by Wang et al. (2021)​​​​​​​ was also used.

Soil erosion data came from the dataset of soil erosion intensity with 300 m resolution over the TP (1992, 2005, 2015) (Zhang, 2019). It was calculated using the China soil loss equation (CSLE) considering precipitation, soil erodibility, slope length, slope, vegetation cover, management, and crop factors. Soil erodibility was also included in the dataset, and the Soil Erodibility Dataset of the Pan-Third Pole in 2020 (Yang and He, 2019​​​​​​​) was used. The data on soil water content were derived from the Global High-Resolution Soil-Water Balance dataset (Trabucco and Zomer, 2010), which defines the fraction of soil water content available for evapotranspiration processes (as a percentage of the maximum soil water content) and is therefore a measure of soil water stress. We calculated catchment-level soil water content at monthly and annual scales. The mean annual ground temperature (MAGT) was derived from the MAGT and permafrost thermal stability dataset over the TP from 2005 to 2015 (Ran and Li, 2019​​​​​​​).

The geological characteristics include lithological class, subsurface permeability, and porosity. The lithological classes came from the Global Lithological Map (GLiM) database V 1.0 (Hartmann and Moosdorf, 2012). GLiM consists of three classification levels, and the first level which contains 16 lithological classes was adopted. The subsurface permeability and porosity, two crucial parameters for groundwater modeling, were derived from the GLobal HYdrogeology MaPS 2.0 (GLHYMPS 2.0) dataset (Huscroft et al., 2018). Permeability measures how easy the rock permits the passage of fluids, and porosity measures how much water can be stored in the subsurface. These two parameters were estimated based on the GLiM lithological map, which can differentiate fine and coarse-grained sediments and sedimentary rocks. To calculate the catchment-level characteristics, the arithmetic mean was used for porosity, while the logarithmic scale geometric mean was used for permeability.

Figure 10 shows the spatial distribution of soil and geology characteristics for inter-lake catchments on the TP. The pH value is high in the west of the Inner TP, and the SOC and total nitrogen content are low in this region. The sand content is high in the south of the Inner TP and the northeast of the TP, while the clay content shows the opposite pattern. The fraction of permafrost extent is high in the north of the Inner TP. The lithological classes show a latitudinal distribution, and the main types include siliciclastic sedimentary rocks, mixed sedimentary rocks, and unconsolidated sediments. The subsurface permeability is higher in the south of the TP than in the north, and the subsurface porosity is higher in the north of the TP and the valley between the Kunlun and Gangdise mountains in the south of the Inner TP.

Population count and density, man-made objects such as cities and roads, and nighttime lights were used to characterize human activities in a catchment. Human footprint, a comprehensive index for evaluating human activities, was also included in this dataset. The population count and density data were obtained from the Gridded Population of the World (GPW) database v4.11 (Center for International Earth Science Information Network – Columbia University, 2018). This database provides estimates of the human population (number of persons per pixel) at a spatial resolution of 30 arcsec for the years 2000, 2005, 2010, 2015, and 2020. Nighttime lights (NLI) are a useful proxy to characterize the intensity of human activity, and the DMSP-OLS Nighttime Lights v4 dataset (Doll, 2008) was used in this study. It was produced using cloud-free remote-sensing images from the Defense Meteorological Satellite Program (DMSP) Operational Linescan System (OLS) at a spatial resolution of 30 arcsec. The values in this dataset represent the product of the average visible band digital number of cloud-free light detections and the percent frequency of light detection.

Road density was derived from the Global Roads Inventory Project (GRIP) dataset (Meijer et al., 2018). Nearly 60 geospatial datasets on road infrastructure (from 1997 to the present) were gathered, harmonized and integrated into a global road dataset. The resulting dataset includes over 21 million km of roads, classified into five types. In this research, catchment-level road density was calculated from a simplified grid dataset at 5 arcmin spatial resolution. Human footprint is a measure of how much we are using the earth's natural resources, and the Global Human Footprint v2 dataset at a spatial resolution of 30 arcsec (Venter et al., 2016) was used. In this dataset, eight different factors, including built environments, population density, electric infrastructure, croplands, pasture lands, roads, railways, and navigable waterways, were combined to measure the direct and indirect human pressures on the environment globally in 1993 and 2009.

Figure 11 shows the spatial distribution of anthropogenic activity characteristics for inter-lake catchments on the TP. The population density, road density, and human footprint all suggest that human activities are relatively intense in the south and northeast of the TP and there is almost no human activity in the north of the Inner TP where elevation is high and environmental conditions are harsh.

This dataset also provides the time series of several important hydrological and meteorological variables (Table S2), including the following: (1) daily meteorological variables (i.e., 2 m air temperature, surface pressure, and specific humidity, 10 m wind speed, downward shortwave radiation, downward longwave radiation, and precipitation amount) from the CMFD dataset covering the period 1979–2018 (Yang and He, 2019); (2) remote-sensed submonthly water level and volume data (2000–2017) extracted from Landsat images and altimetry data based on lake shoreline positions (Li et al., 2019), ​​​​​​​∼ monthly water level data (2010–2020) extracted from multi-sensor altimetry data (Xu et al., 2022), lake area and mass change data at 5-year intervals (1976–2020) extracted from satellite stereo and multispectral images (Zhang et al., 2021); (3) remote-sensed daily fractional snow cover based on the MODIS surface reflectance product MO/YD09GA covering the period 2000–2022 (Jiang et al., 2022), daily snow depth data (1980–2019) produced through the fusion of five gridded snow depth datasets using machine learning methods (Che et al., 2021), and daily snow water equivalent data (2002–2011) based on AMSR-E brightness temperature (Qiu, 2018b); (4) yearly glacier mass change rates (2000–2019) extracted from large-scale and openly available satellite and airborne elevation datasets (Hugonnet et al., 2021); and (5) decadal maximum freezing depth data of seasonal frozen soil (1961–2020) produced by the support vector regression model based on in situ measurements from 2001 to 2010 and spatial environmental variables (Wang and Ran, 2021). These time series data facilitate the analysis of temporal variation at the catchment scale and can be used for hydrological modeling based on lumped hydrological models or machine learning methods.

The RMSE of the Landsat-derived water levels from Li et al. (2019) was 0.11 m. The water level data from Xu et al. (2022) had R2>0.80 and RMSE <0.12 m in Qinghai Lake. The uncertainties for each value in the time series of Li et al. (2019), Zhang et al. (2021), and Xu et al. (2022) can be found in the corresponding uncertainty files (Table S2).

Most topographic attributes in this dataset were derived from MERIT DEM and MERIT Hydro (flow direction map) datasets. MERIT DEM was produced by eliminating main error components (e.g., absolute bias, stripe noise, speckle noise, and tree height bias) from existing DEMs (SRTM3 DEM, AW3D DEM, and VFP-DEM). It has a resolution of 3′′ (∼ 90 m at the Equator) and the land areas mapped with ±2 m or better vertical accuracy were 58 % (Yamazaki et al., 2017). MERIT Hydro was derived from MERIT DEM and water body datasets (G1WBM, Global Surface Water Occurrence, and OpenStreetMap). The relative error of MERIT Hydro in drainage area delineation was less than 0.05 for 90 % of Global Runoff Data Center (GRDC) gauges.

The CMFD meteorological dataset used in this study was produced through fusion of remote-sensing products, reanalysis datasets, and in situ observations from a larger number of stations. Its accuracy in western China was validated based on independent observations, and the results showed that CMFD had closer-to-zero mean bias error (MBE), lower RMSE, and higher R2 than the Global Land Data Assimilation System (GLDAS) for almost all meteorological variables (He et al., 2020).

The land cover/use data used in this study came from the fusion of six popular land use products, with an accuracy of 88.71 % (Xu, 2019). The GPP and NPP data came from the MODIS products (MOD17A2H.006 and MOD17A3HGF.006). The R2 between monthly MODIS GPP and eddy covariance measurements was reported to be 0.64 on average, and the RMSE was 2.55 g C m-2 d-1 in alpine grassland, which is the most widely distributed biome on the TP (Zhu et al., 2018); the R2 between MODIS NPP and in situ observations in 23 stations across China was reported to be 0.81, and the RMSE was 73.44 g C m-2 (Sun et al., 2021). The RMSE of fractional snow cover data from Jiang et al. (2022) was 0.14 taking the results from high-resolution Landsat images as reference. The R2 between snow depth data from Che et al. (2021) and in situ observations was 0.81, and the RMSE and mean absolute error (MAE) were 7.7 and 2.7 cm, respectively.

The SoilGrids 2.0 dataset used in this study was generated by machine learning methods, using approximately 240 000 soil observations worldwide and over 400 environmental variables as inputs. It provides a spatial distribution map of data uncertainty generated by the quantile regression forest prediction model, which is the ratio of the interquartile range (i.e., the difference between 0.95 quantile and 0.05 quantile) over the median (Poggio et al., 2021). The catchment-level average uncertainty for each soil variable was calculated and included in this dataset. For the maximum freezing depth of seasonal frozen soil, the R2 in the four periods of 1980s, 1990s, 2000s, and 2010s were 0.77, 0.83, 0.73, and 0.71, respectively (Wang and Ran, 2021).