The major deserts in China include the Taklamakan Desert, Gurban Tunggut Desert, Qaidam Desert, Kumtag Desert, Badain Jaran Desert, Tengger Desert, Ulan Buh Desert, Hobq Desert, MU US Desert, Hunshandake Desert, Hulunbuir Sands, and Horqin Sands. All the desert boundaries were derived from Google Earth Pro® via manual interpretation. We delineated the desert boundaries using the Digital Global Feature Imagery and SpotImage (2011, 10 m resolution) collections of Google Earth Pro®, whose spatial resolution is finer than 30 m. The acquisition time of most images was in 2011.

Gf-2 satellite is the first civil optical remote sensing satellite independently developed by China with a spatial resolution better than 1 meter. It is equipped with two high-resolution 1-meter panchromatic and 4-meter multi-spectral cameras, and the spatial resolution of the sub-satellite can reach 0.8 meters. This data set is the remote sensing image data of 6 jing gaofen-2 satellite in 2017.The folder list is: GF2_PMS1_E100.5_N37.2_20171013_L1A0002678101 GF2_PMS1_E100.5_N37.4_20171013_L1A0002678097 GF2_PMS1_E100.6_N37.6_20171013_L1A0002678096 GF2_PMS2_E100.3_N37.4_20170810_L1A0002534662 File naming rules: satellite name _ sensor name _ center longitude _ center latitude _ imaging time _L****

This data set is the remote sensing data of gaofan-1 satellite, including the data of two scenes of PMS1 camera on 2017-8-13 and 2017-10-5, one scene of PMS2 camera on 2017-5-27, and one scene of WFV2 and WFV3 camera on September 23, 2018.File list: GF1_PMS1_E99.1_N37.2_20170813_L1A0002539236 GF1_PMS1_E101.2_N36.4_20171005_L1A0002653985 GF1_PMS2_E100.3_N37.7_20170527_L1A0002384098 GF1_WFV2_E98.4_N37.6_20180927_L1A0003481737 GF1_WFV3_E100.4_N37.3_20180927_L1A0003481706

The data set is extracted from the NDVI data of long time series acquired by VEGETATION sensor on SPOT satellite. The time range of the data set is from May 1998 to 2013. In order to remove the noise in NDVI data, the maximum synthesis is carried out. A NDVI image is synthesized every 10 days. The data set is cut out from the global data set, so as to carry out the research and analysis of the source areas of the three rivers separately. The data format of this data set is geotiff, spatial resolution is 1 km, temporal resolution is 10 days, time range: May 1998 to December 2013.

The data set is NDVI data of long time series acquired by NOAA's Advanced Very High Resolution Radiometer (AVHRR) sensor. The time range of the data set is from 1982 to 2015. In order to remove the noise in NDVI data, maximum synthesis and multi-sensor contrast correction are carried out. A NDVI image is synthesized every half month. The data set is widely used in the analysis of long-term vegetation change trend. The data set is cut out from the global data set, so as to carry out the research and analysis of the source areas of the three rivers separately. The data format of this data set is GeoTIFF with spatial resolution of 8 km and temporal resolution of 2 weeks, ranging from 1982 to 2015. Data transfer coefficient is 10000, NDVI = ND/10000.

The data set is MODIS vegetation index data (MOD13Q1). The source areas of the three rivers are extracted to carry out the research and analysis of the source areas of the three rivers separately. MOD13Q1 is a 16-day composite vegetation index, including normalized vegetation index (NDVI) and enhanced vegetation index (EVI). The spatial scope of Sanjiang Source covers two MODIS files (h25v05 and h26v05). Data storage format is hdf. Each file contains 12 bands: Normalized Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Data Quality (VI Quality), Red Reflectance, Near Infrared Reflectance (NIR Reflectance), Blue Reflectance, Mid Infrared Reflectance, Observation. Viewzenith angle, sun zenith angle, relative azimuth angle, composite day of the year and pixel reliability. The data format of this data set is hdf, spatial resolution is 250m, temporal resolution is 16 days, time range: February 2000 to October 2018.

This is the vegetation index (NDVI) for Maduo County in July, August and September of 2016. It is obtained through calculation based on the multispectral data of GF-1. The spatial resolution is 16 m. The GF-1 data are processed by mosaicking, projection coordinating, data subsetting and other methods. The maximum synthesis is then conducted every month in July, August, and September.

The data set is based on the reflectance of MODIS channels and the observation data of SIF to establish the neural network model, so as to obtain the SIF data with high spatial and temporal resolution, which is often used as a reference for primary productivity. The data is from Zhang et al. (2018), and the specific algorithm is shown in the article. The source data range is global, and the Tibetan plateau region is selected in this data set. This data integrates the original 4-day time scale data into the monthly data. The processing method is to take the maximum value of the month, so as to achieve the effect of removing noise as much as possible. This data set is often used to evaluate the temporal and spatial patterns of vegetation greenness and primary productivity, which has practical significance and theoretical value.

The dataset is the ground verification point dataset of land cover and vegetation type in the Source Region of Yellow River (in the north of Zaling Lake, Qinghai Province) which collected during August 2018. In the dataset, the homogeneous patches are considered as the main targets of this collection. They are easy to be recognized out and distinguished from other vegetation types. And these samples have high representativeness comparing with other land surface features. In each sample, the geographical references, longitude and latitude (degree, minute, second), time (24h) and elevation (0.1m) are recorded firstly according to GPS positioning. Vegetation types, constructive species, characteristics, land types and features, landmarks, etc. are recorded into the property table manually for checking in laboratory. At last, each sample place has been taken at least 1 photography. In this dataset, 90% or more samples have been taken 2 or more in field landscape photographs for land use type and vegetation classification examination. We have carefully examined the position accuracy of each sample in Google Earth. After 2 rounds of checking and examination, the accuracy and reliability of the property of each sample have been guaranteed.

The dataset is the ground verification point dataset of land cover and vegetation type in the Hoh Xil (in the northwest of Qinghai Province) which collected during August 2018. In the dataset, the homogeneous patches are considered as the main targets of this collection. They are easy to be recognized out and distinguished from other vegetation types. And these samples have high representativeness comparing with other land surface features. In each sample, the geographical references, longitude and latitude (degree, minute, second), time (24h) and elevation (0.1m) are recorded firstly according to GPS positioning. Vegetation types, constructive species, characteristics, land types and features, landmarks, etc. are recorded into the property table manually for checking in laboratory. At last, each sample place has been taken at least 1 photography. In this dataset, 90% or more samples have been taken 2 or more in field landscape photographs for land use type and vegetation classification examination. We have carefully examined the position accuracy of each sample in Google Earth. After 2 rounds of checking and examination, the accuracy and reliability of the property of each sample have been guaranteed.

The dataset is the ground verification point dataset of land cover and vegetation type in the Source Region of the Yangtze River (in the south of Qinghai Province) which collected during August 2018. In the dataset, the homogeneous patches are considered as the main targets of this collection. They are easy to be recognized out and distinguished from other vegetation types. And these samples have high representativeness comparing with other land surface features. In each sample, the geographical references, longitude and latitude (degree, minute, second), time (24h) and elevation (0.1m) are recorded firstly according to GPS positioning. Vegetation types, constructive species, characteristics, land types and features, landmarks, etc. are recorded into the property table manually for checking in laboratory. At last, each sample place has been taken at least 1 photography. In this dataset, 90% or more samples have been taken 2 or more in field landscape photographs for land use type and vegetation classification examination. We have carefully examined the position accuracy of each sample in Google Earth. After 2 rounds of checking and examination, the accuracy and reliability of the property of each sample have been guaranteed.

The permafrost stability map was created based on the classification system proposed by Guodong Cheng (1984), which mainly depended on the inter-annual variation of deep soil temperature. By using the geographical weighted regression method, many auxiliary data was fusion in the map, such as average soil temperature, snow cover days, GLASS LAI, soil texture and organic from SoilGrids250, soil moisture products from CLDAS of CMA, and FY2/EMSIP precipitation products. The permafrost stability data spatial resolution is 1km and represents the status around 2010. The following table is the permafrost stability classification system. The data format is Arcgis Raster.