Dune movement under climatic changes on the north‐eastern Tibetan Plateau as recorded by long‐term satellite observation versus ERA‐5 reanalysis

The movement of active dunes is tightly linked to climatic conditions (e.g., wind regime, temperature and precipitation) as well as human influence (e.g., grazing, dune fixation and greening). Dune migration rates can be studied to draw conclusions of changing wind conditions over time. The Gonghe Basin (GB), located on the north‐eastern Tibetan Plateau (TP), offers a good testing ground for these assumptions. The intramontane basin is highly influenced by two major wind regimes: the mid‐latitude Westerlies and the East Asian summer monsoon. To investigate environmental changes, this study combines optical remote sensing techniques with climatic datasets. High‐resolution satellite images of the last five decades, such as CORONA KH‐4B, are used to map dunes and calculate their respective migration rates. Further, height information was extracted as well. Climatic changes from the ERA‐5 reanalysis dataset and normalized difference vegetation index (NDVI) values were processed alongside. Relating the dunes' surface processes to climate model data shows an accordance between slowing migration, expanding vegetation and a decrease in sand drift potential. From 1968 to present time, an average dune migration rate of 7.3 m a−1 was extracted from the satellite images, with an overall reduction of −1.81 m a−1. The resultant drift potential (RDP) values for the GB are calculated to be below 10 m3 s−3 with a spatial decrease, following a direction from the NW to the SE, fitting well with a corresponding decrease in the migration rates. Our results indicate a good agreement between the development of aeolian landforms and the ERA‐5 climate reanalysis model data, even in a high‐altitude setting with complex topography, which is known to influence such datasets.


| INTRODUCTION
Dryland areas cover about 40% of the land surface and are home to almost the same percentage of the world's population (Reynolds et al., 2007;White & Nackoney, 2003).In addition, they are highly sensitive to climatic changes (Huang et al., 2016(Huang et al., , 2017)), making it increasingly important to understand the linkages between surface processes in arid and semi-arid regions and climate change (Huang et al., 2016(Huang et al., , 2019)).Global expansion of drylands was reported for the last 60 years (Feng & Fu, 2013;Huang et al., 2016).However, regional differences occur and need to be examined further (Cuo et al., 2013;Xu et al., 2008).The northern and north-eastern Tibetan Plateau (TP) lies mainly within the influence of the East Asian summer monsoon (EASM) (Wang et al., 2005) and the mid-latitude Westerlies (An et al., 2012), resulting in seasonal changes of the prevailing wind directions.The orographic effect from the TP causes widespread arid and semi-arid conditions (Kong & Chiang, 2020).Consequently, aeolian sand sheets and dune systems are abundant on the north-eastern TP (Stauch et al., 2018).Due to its location between these diverse climate systems, climatic changes show a multitude of effects on the TP (Chen et al., 2021;Cuo et al., 2020;Zhang et al., 2017Zhang et al., , 2021)).
In recent years, a warming and wetting trend is reported by a number of researchers for the central and northern TP (Chen et al., 2021;Gerlitz et al., 2014;You et al., 2010) and was also linked to the topography of the TP (Qin et al., 2009).The warming trend is reflected in the melting of snow and ice coverage, the retreat of glaciers and permafrost reduction (Li & Sheng, 2012;Yao et al., 2007Yao et al., , 2012;;Ye et al., 2006).This results in a decrease of albedo, which in turn favours further glacier melting (Pekel et al., 2016;Zhang et al., 2021;Zhang, Hu, et al., 2020) as well as rising lake levels and subsequent grassland flooding (Pekel et al., 2016;Zhang, Yao, et al., 2020).In general, most observations show a weakened northern extension of the EASM in the second half of the 20th century (Zhu et al., 2012), as well as a northward shift of the monsoon circulation (Dong et al., 2021;Goldsmith et al., 2017;Sooraj et al., 2015;Wang et al., 2022).Frequently, studies assess monsoonal patterns by using climate models like ERA-5, JRA-55 or MERRA-2 (Ceglar et al., 2017;Hamal et al., 2020;Liu & Margulis, 2019;Quagraine et al., 2020).In areas, where the number of climate stations is low and climate records are comparably short, like the TP (Curio & Scherer, 2016;Gerlitz et al., 2014;Klinge et al., 2003), worldwide climate models, such as the ERA-5 dataset, can fill the gaps in observational data.The steep and complex topography of the TP, however, was found to have an influence on climate models, when compared to the areas of lower elevation surrounding the TP (Fu et al., 2021).This often results in a wet bias, when compared to local precipitation observations (Gao et al., 2015;Meehl & Arblaster, 1998;Phillips & Gleckler, 2006;Ying & Chong-Hai, 2012).A similar effect on the climate models from complex topography was found by Velikou et al. (2022) for the European Alps.
A possible way to evaluate the performance of climate models is the analyses of geomorphological landforms which are showing quick responses to climatic changes.Dunes are a common landform in semiarid environments (Tsoar, 2001).They react quickly to changes in environmental conditions and were used to reconstruct long-and short-term climatic changes (Forman et al., 1992;Gaylord & Stetler, 1994;Lancaster, 1988;Muhs, 1985;Vimpere et al., 2021).
The exact threshold speed depends on the size and shape of the sand grains, as well as the surrounding surface conditions influencing the aerodynamic roughness length (Valance et al., 2015).Wind-driven erosion is one of the most influential factors on the migration of unfixed dunes.However, moisture and resulting vegetation are also able to demobilize dunes (Stauch et al., 2017(Stauch et al., , 2018)).In order for vegetation to grow naturally on dune bodies, the wind speed needs to drop significantly.Also, to remobilize vegetated dunes, the wind has to rise to high speeds again to overcome fixation (Tsoar, 2005).Therefore, vegetation development and the resulting slowing of dune migration can be viewed as indications of climatic changes.Another factor influencing dunes is human activity.Fixation through planting checkerboard pattern vegetation, for example, is a common measure to prevent dunes from migration through settled areas and across existing infrastructure (Qi et al., 2021).Conversely, desertification and subsequent dune migration can be enhanced or even induced by overgrazing through poorly managed livestock and trampling of bare soil through animals (Bauer et al., 2011;Bauer & Nyima, 2011;Lehmkuhl et al., 2003;Wang et al., 2018).
Remote sensing data have been used to study the migration and development of dunes for decades (Finkel, 1959).The ongoing improvement of temporal, spectral and spatial resolution of remote sensing data has pushed the study of dunes and their migration since the first satellite images were available and the technical requirements were met in the 1970s (Forman et al., 1992;Hugenholtz et al., 2012).
Freely available remote sensing images from the Landsat and Sentinel satellite programs provide a continuous database with a high spatial coverage.Going further back in time, satellite images of sufficient resolution become sparser.In the 1990s, the American government declassified images from espionage satellites of the 1960s CORONA

program (Earth Resources Observation and Science [EROS]
Center, 2017a).A multitude of film containers were retrieved successfully, resulting in a big database of high-resolution monochrome satellite images covering 35 million km 2 of the Earth's surface (Scollar et al., 2016).
The Gonghe Basin (GB) on the northern TP Plateau is influenced by the EASM and the mid-latitude Westerlies.Due to its semi-arid conditions and the widespread occurrence of sand, active dunes and sand sheets are common in the GB (Stauch et al., 2017(Stauch et al., , 2018)), which is the location for observations of this study.Climate station is rare in the area, and measurements are often not continuous (Böhner, 2006).
Aeolian landforms might be suitable indicator for climatic changes in this remote environment.Furthermore, the complex topography at the northern edge on the TP makes it a suitable area to compare the evolution of aeolian landforms and ERA-5 wind data under climate change.
In the presented study, we explore the potential of dunes as highly reactive landscape features to fill the gap that prevails due to missing observations in high-altitude regions featuring complex topography.We derive dune migration rates by tracking their respective positions through time, utilizing satellite images.The high-resolution images are also used to estimate their heights.The ERA-5 datasets are used to calculate sand drift potential (DP) and to derive individual RDP values for each single dune, as well as to evaluate precipitation and temperature values for the past five decades.Concomitantly, multitemporal satellite images are applied to evaluate vegetation cover through the normalized difference vegetation index (NDVI) in a time series.

| Study area
With a size of 20 000 km 2 , the GB (Figure 1) (36 N, 100.5 E) is the largest intramontane basin on the north-eastern TP.It stretches for about 250 km in a NW-SE direction.The widest part is roughly 60 km.The geological formation of the GB started around 14.6 myr ago, and the uplift of the neighbouring mountains is still ongoing (Lu et al., 2012;Zhang et al., 2012).The GB is crossed by several faults (Xu et al., 2019).The last major earthquake occurred in 1990 with a magnitude of 6.9 (e.g., Hao et al., 2012).Neogene and Quaternary fluvial gravels and lacustrine sediments with a thickness of several hundreds of meters are filling the central part of the basin (Craddock et al., 2010;Perrineau et al., 2011).Based on hydrological criteria, the basin can be subdivided into four different catchments.The northwestern part is dominated by the Chaka Salt Lake (Figure 1), while the central part consists of several small endorheic catchments.The two south-eastern parts are connected to the Yellow River (Stauch, 2016).
The Yellow River, the second longest river in China with a length of 4845 km, is artificially dammed at the south-eastern part of the basin, resulting in a 296 km 2 large lake (Figure 1).Incision of the Yellow River of presently up to 500 m started sometimes between 500 kyr (Craddock et al., 2010) and 150 kyr before present (Perrineau et al., 2011).The Longyangxia Gorge Reservoir was constructed in 1989 and had a profound influence on the ecology but especially on the socio-economic development of the area.Strongly increasing life stock amounts since the mid of the 20th century and an attempt to cultivate farmland in the semi-arid basin in the late 1980s resulted in severe desertification processes (Yan et al., 2002(Yan et al., , 2009)).
The main climatic influence in the GB is related to the wind systems of the EASM and the mid-latitude Westerlies (Qiang et al., 2013).The EASM is characterized by seasonal reversals in its atmospheric circulation and associated precipitation, resulting in wet summers and dry winters (Wang et al., 2014).A moisture front is resulting from the interplay of dry and cold air masses, transported by the westerlies (Figure 1) and warm and wet air from the EASM (Molnar et al., 2010).
The local present-day climate of the GB is characterized by a mean annual air temperature of 3.7 C and a mean annual precipitation of $310 mm based on stationary measurements , with 80% of the precipitation occurring between May and September (Qiang et al., 2013).In contrast, the ERA-5 dataset yields a higher estimate mean annual precipitation of $480 mm for the past five decades ).An arid climate has been dominating the basin during the early Holocene followed by a moisture optimum in the middle Holocene and a drying trend afterwards (Liu et al., 2020;Stauch et al., 2018).Nowadays, strong winds from north-western and northern directions are conveying cold and dry air masses to the study area during winter and spring (Qiang et al., 2013).(Dashora et al., 2007).Most commonly, these images are used for landscape analysis and change detection (Bayram et al., 2004;Fekete, 2020;Mihai et al., 2016;Poli et al., 2012;Song et al., 2015;Stauch et al., 2014;Stauch & Lehmkuhl, 2010).However, others have utilized the stereoscopic camera alignment (Galiatsatos et al., 2005) to create digital elevation models for landscape change analysis (Jacobsen, 2020;Meszaros et al., 2008;Mihai et al., 2016;Schmidt et al., 2001).The images used for this work are from the KH-4B missions 1102, 1105 and 1112 from the year 1968.As the KH-4B images were taken with an analog camera and scanned on demand, they are not referenced to any coordinate system.Further affected by the camera arrangement inside the satellite, this results in fairly high optical distortions (Scollar et al., 2016) For the NDVI raster of the GB, the red and near infrared bands of a Copernicus Sentinel-2 scene of August 2021 were used.

| Dune mapping
Mapping of the dunes was done manually.For the migration rates, mostly singular dunes were chosen.
To detect changes within dune migration rates, measurements from all time periods were added.The images chosen result in time intervals of 19, 14 and 19 years from for T1, T2 and T3, respectively.
Quite typically, slip faces, crestlines or vegetation marks are selected to represent the dunes position (Jimenez et al., 1999;Levin et al., 2009).Crestlines were chosen to be mapped instead of the full barchan dune bodies, since they are well recognizable on the older and newer satellite images and even on the Landsat imagery with a coarser spatial resolution.This is due to their high contrast between the windward-oriented-dune body and high reflection of the sunlit slip face (Baughman et al., 2018).Due to the differences between the

| Dune height estimations
A simple trigonometrical approach was used to extract dune heights from the satellite images based on the idealized shape of a barchan cross section.By measuring the horizontal base line b from the midpoint of the crestline to the closest toe of the slip face, the height h of the barchan can be calculated trigonometrically given the angle of repose α is known: For α, a value of 31 ± 1 is assumed.As of literature research and observed values from field measurements, angles between 28 and 33 are common for most barchan dunes (Ahnert, 1996;Bagnold, 1954;Zhu et al., 2021).Furthermore, we assume the midpoint of the crestlines to be at least close to the highest point.For symmetrical shaped barchans, this can be considered a valid assumption (Hesp & Hastings, 1998;Kroy et al., 2005).
As a measure of uncertainty for the dune height estimation, we assessed the error for the baseline by propagating the uncertainty at each measurement point.This measurement is taken from the resolution of the images from each time point.Due to differing image resolutions, the uncertainties result in different upper and lower values (Table 2).Given the above chosen angle of repose α of 31 ± 1 , maximum and minimum heights were subsequently calculated for every dune.
In contrast, a construction of DEMs based on the stereoscopic setup of the CORONA images and subsequent dune height derivation is not considered feasible as many geometric image parameters are unknown and their estimation would result in large uncertainties (Meszaros et al., 2008;Schmidt et al., 2001).et al., 2021;Hassler & Lauer, 2021;Lei et al., 2022;Velikou et al., 2022;Xu et al., 2022).ERA-5 shows, in comparison with CFS-2, MERRA-2 and JRA-55, a good overall representation of precipitation intensities and extreme events, whereas JRA-55 shows an overestimation in low-annual-precipitation regions (Arshad et al., 2021).
The same was found for multi-year monthly precipitation (Arshad et al., 2021).In summary, ERA-5 shows an overall better reproduction of precipitation data (Hassler & Lauer, 2021).
The Google Earth Engine (GEE) (Amani et al., 2020;Zhao et al., 2021) was used to enable an in-depth multitemporal analysis of the ERA-5 climate dataset (wind, precipitation and temperature) and of the NDVI.
For the multitemporal NDVI analysis, satellites of the Landsat mission were used to ensure the longest possible timespan.To achieve a single continuous image collection, pre-processed surface reflectance products of Landsat 5, 7 and 8 were used.The near infrared and red bands from the Landsat satellites are pre-processed and atmospherically corrected with the Landsat Ecosystem Disturbance Adaptive Processing System (LEDAPS) and the Land Surface Reflectence Code (LaSRC) algorithms for Landsat 5 and 7 and for Landsat 8, respectively (Masek et al., 2006).This continuous collection includes more than 5500 scenes, dated between June 1986 and December 2021.Before the calculation of the NDVI, clouds were masked based on the pixel QA band.The NDVI itself was calculated as a normalized difference between the NIR and red spectral bands.
The NDVI was calculated using the following formula (Rouse & Haas, 1974;Tucker, 1979): The NDVI was calculated for the years 1987 to 2019, due to the full coverage by satellite images in this years.For the seasonal analyses, the NDVI period was split into two periods equal to T2 (1987-2001) and T3 (2001-2019).
The following method to calculate these sand drift regime parameters was adapted from Baas and Delobel (2022).Hourly 10 m u-and v-components of the wind were downloaded from the ERA-5 dataset for the whole GB and for three specific point locations (36.37 N, 99.8 E; 36.15N, 100.4 E; 35.9 N, 100.5 E).The daily average scalar magnitude of DP was derived from the 24-hourly wind speed values (computed from the u-and v-component of the wind) following Equation (3): where U is the surface wind speed at 10 m above ground (m s À1 ) and U t is the wind speed threshold for initiating sand transport, set here at 6 m s À1 (Fryberger, 1979;Yang et al., 2021).From the daily scalar magnitude of DP and the u-and v-components of the wind, the vector direction of DP was calculated and averaged for each day.The overall DP for each grid cell of the GB or point location was computed from the mean of all daily-averaged DP magnitudes whereas the RDP vector was calculated from the vector sum of all daily-average DP vectors.The RDDir was computed from the RDP vector, and RDP/DP was calculated from the ratio of RDP vector over the overall DP.
The DP, RDP, RDDir and RDP/DP values were all averaged for the three areas and for all three time periods T1, T2 and T3.DP and RDP values are given in m 3 s À3 .The seasonal sand roses for the three specific locations were created in MATLAB by compiling all the hourly 10 m u-and v-components of the wind per season and then following the method described in Baas and Delobel (2022).

| RESULTS
To decipher the climate influence on dune migration in the GB, we first present our results of the dune mapping based on different satellite images and then evaluate these in the light of the ERA-5 reanalysis climate data.

| Dune measurements, types and distribution
The predominant dune type in the GB is the barchan.Various longitudinal dunes have developed as well.Most freely moving barchans in the GB can be found north-west of the Longyangxia reservoir.On the western shores of the reservoir, the barchans have collided and formed contiguous dune fields, and the reservoir acts as a barrier for further migration.South-east of the reservoir, hardly any singularly moving dunes are found.Here, a big dune field is located, fixed with checkerboard straw patterns.Consequently, this dune field was excluded, and 76 individual dunes were chosen for further analysis.
All dunes are located north-west of the Longyangxia reservoir.
The dunes are mapped in three main areas (C1, C2 and C3; Figure 1).
• Cluster C1: the north-western-most dunes, a small group of about 30 barchans, located directly along the Shazhuyu River.
• Cluster C2: the main dune field in the middle of the basin roughly 30 km downwind from the previous group, comprising singular and coupled active barchans.
• Cluster C3: large barchans or transverse dunes within the continuous dune fields at the western shore of the Longyangxia water.

| Barchan dune migration
For the time interval of 52 years (1968-present), an average dune migration rate of 7.3 ± 0.04 m a À1 was calculated (Table 1).
Fifty-two of the 76 dunes move in a south-eastern direction, a further 17 in the eastern direction, following the mean wind direction (Figure 3).The remaining seven dunes are migrating in a southern or south-western direction.These seven dunes are located within the dense dune field at the western shore of the reservoir.

| Dune height estimates
Dune height estimates in the GB vary between 2 and 6 m and 40-45 m in 1968 and 2-4 m to 38-39 m now, resulting in an average decrease of 17% in dune height (Table 2; Figure 4).The range of heights and the heights of the highest and the lowest dunes have also decreased.S1).

| NDVI
T A B L E 1 Dune migration lengths and rates from 1968 to 2019.

| ERA-5 precipitation and temperature
From the ERA-5 dataset, daily and monthly temperature and precipitation values were obtained (Figure 6).The measured mean annual precipitation for the GB is 486 mm, roughly 176 mm higher than the reported mean based on stationary measurements between 1957 and 2000 (Qiang et al., 2013).The annual mean air temperature was found to be lower in the ERA-5 data, with 2.15 C, instead of the mean 3.7 C reported by Qiang et al. (2013).For both ERA-5 climate parameters, a slight increase over the time span of the past 35 years is found (Figure 6).

| Sand drift potential
The DP and corresponding metrics were calculated for three clusters (Table 3).RDDir values are in radian.The DP values for all three locations are below 10 m 3 s À3 , which corresponds to a low sand drift environment (little activity), but there is a decrease in DP from the fastest barchans (higher activity/migration) to the transverse dunes (lower activity) (this decrease in activity/DP is also visible in the sand roses).There are very little differences between the DP values in the grid cells compared to the point locations.The high RDP/DP values (close to 1) presented here indicate a unidirectional sand drift regime (Fryberger, 1979;Lancaster, 2009), which corresponds well to the barchan and transverse dunes seen in the satellite imagery.
When comparing the drift potential to the migration rates, we find a strong agreement in their development (Figure 7).A Pearson's correlation further showed a significant positive correlation (Figure 7).
A significant correlation between migration rates and their respective height estimates was not found.For RDP value extraction, a point grid of RDP values was created with a 100 m spacing between the points.
This point grid was than interpolated to a raster set with a 1000 m resolution in order to pin single RDP values to each individual dune.

| DISCUSSION
The study area of the GB with its location on the TP is influenced by two different wind systems, the EASM and the mid-latitude Westerlies.The general wind directions are well represented in ERA-5 climate model.However, the complex topography of the TP has proven influential on the models (Gao et al., 2015;Meehl & Arblaster, 1998;Phillips & Gleckler, 2006;Ying & Chong-Hai, 2012).
Seventy-six dunes in the GB were identified and mapped on satellite images from 1968, 1987, 2001 and present time in this study.
On a worldwide scale, dune migration rates vary quite a lot, depending on their respective location, vegetation cover, sand availability and climatic situation, like the prevailing wind system or precipitation at their location.However, besides climatic parameters, also topography influences wind speeds and therefore the dune migration speed (Iversen & Rasmussen, 1999).This can be reflected by wind-channelling effects, common for narrow sections in mountain basins like the GB (Whiteman, 2000).The overall migration rates of dunes in the GB, compared globally, ranked rather slow with 7.3 ± 0.04 m a À1 .
Slow dunes, for instance, are migrating with rates under 5 m a À1 in Wales and Alaska due to high precipitation and vegetation cover and, at the Namib Sand Sea, with low sand supply in the interdune area (Bailey & Bristow, 2004;Bristow & Lancaster, 2004;Necsoiu et al., 2009).On the other hand, high rates have been reported from the Atacama Desert with a migration rate of 60 m a À1 where vegetation is mostly absent and strong unidirectional winds are prevailing (Parker Gay, 1999) and even 95.2 m a À1 in the Sanlongsha dune field in China, being an arid environment with sparse vegetation cover and strong seasonal winds (Yang et al., 2021).The fastest dunes found in    outside of the field.These are, however, the rates for the whole time span.Looking at the three time periods, we find a decreasing trend overall and faster rates in the first time span (9.50 ± 0.15 m a À1 ).
In the dune fields near the reservoir (C3), the rates have dropped from 2.36 ± 0.15 m a À1 for the first period to 1.94 ± 0.37 m a À1 in the last.In the second period, we find an increase in the migration rates in areas C1 and C3, contrasting a steady decrease in C2.
This slight rise, however, is not found in the RDP averages for all three locations, potentially due to resolution uncertainty or smallscale effects on the studied dunes.In the third period, all three areas have lower rates than in the first period, confirming the overall decrease.The migration direction shows a clear majority of 68% of the mapped dunes to migrate in a south-eastern direction, which is in accordance with the prevailing winds coming from the north-west, as calculated from ERA-5 (Figure 3).Here, the calculated resultant drift direction (RDD) ranges from 104 to 131 (average 110 ) angle units (Table 3).The measured directions from mapping are ranging between 96 and 209 (average 127 ), further confirming the findings from the climate models.
For the whole basin, the ERA-5 analysis shows hardly any change in the already low DP.The same goes for the RDP and the RDP/DP.The RDP raster was calculated for a grid, which includes the Qinghai Lake, which is not part of the GB.Here, we find the higher RDP values of up to 12 m 3 s À3 .Restricted to the GB, the highest values are 8.57, 6.51 and 7.82 m 3 s À3 for periods T1, T2 and T3, respectively.Their corresponding means are as low as 2.95 (T1), 2.09 (T2) and 2.74 (T3) m 3 s À3 , clearly showing a low RDP environment.However, it must be kept in mind that those values include all seasons, and therefore, the seasonal minima during summer and autumn are incorporated in these results.The Fryberger classification of wind environments (Fryberger, 1979) was developed for calculations with knots.As we are using metrical values, these scales are not fully comparable (Bullard, 1997).According to Baas and Delobel (2022), high sand drift environments are therefore characterized by DP values >50 m 3 s À3 and low sand drift environments by <10 m 3 s À3 .For all three selected locations of interest (Figure 1), the DP has been averaged seasonally for all three time periods (1968-1987, 1987-2001 and 2001-2019)  wind regime (Fryberger, 1979;Tsoar, 2005).As crescentic barchan dunes can be expected where RDP/DP exceeds 0.7 (Lancaster, 2009), the range assessed for the GB is in total agreement with our observations by mapping based on the satellite images.
The NDVI values show an increase in all seasons for all time periods, indicating rising fractional vegetation cover, as they are highly correlated (Carlson & Ripley, 1997).Vegetation cover is known to act as an element of higher roughness, transferring part of the winds momentum to the roots (Durán & Herrmann, 2006).In addition, vegetation can act as a sand binding agent, weather naturally emerging or artificially planted (Li et al., 2014;Waldron, 1977), resulting in less sediments available to build up dunes.As the vegetation develops first at the horns (Lee et al., 2019), crests and surroundings of existing dunes (Durán & Herrmann, 2006), the body area of the dune will still be subject to erosion, consequently reducing the height of existing dunes, which agrees with our findings.Typical climatic changes described as initiators of vegetation establishment are increasing annual precipitation and the reduction of wind strength (Anthonsen et al., 1996;Gaylord & Stetler, 1994), both of which are recorded in the ERA-5 models for our research area.Therefore, the ERA-5 data and wind drift in particular correlate well with the dune migration rates.Consequently, the ERA-5 data can be used as a good approximation of climatic changes in the GB, at least when applied to the aeolian regime.Also, for the second half of the century, predictions show a further upcoming decrease in Sand Drift Potential for the TP (Baas & Delobel, 2022).Human influence is an important additional factor influencing dune morphology and dune migration.The GB was and is subject to major human influence on the dune landscape (Qi et al., 2021;Wu et al., 2020).The rising temperatures and precipitation have resulted in an increase of grassland converted to agriculture (Wang et al., 2018) and irrigated land (Yongnian et al., 2003).By mapping the agricultural areas on the satellite images from 1968 and 2022, a rise in the agricultural area of 23.7% (from 636 to 787 km 2 ) can be found.
Furthermore, it has been observed, that the construction and recent enlargement of the Gonghe Photovoltaic Power Plant have impacted the local climate, for instance, an increase in daytime temperatures and the nearby humidity, resulting in conditions favourable for vegetation build-up (Wu et al., 2020).This can result in a raised viscosity of sand grains to being less easily entrained by wind, potentially reducing the frequency of sandstorms (Wu et al., 2020).In addition, impacts of human infrastructure on former dune fields can be observed, when comparing satellite images prior to the construction of the Longyangxia reservoir in 1992 with images after.Although these human influences are evident in observations, a direct relation to our collected data was not found.

| CONCLUSIONS
In this study, a multitude of satellite remote sensing data was used to obtain dune migration rates and height estimates in the GB from a period of 52 years-a longer period than in most remote sensingbased dune migration studies.These were compared to climatic developments.The GB is especially prone to climatic changes, due to its location on the TP and seasonally differing wind regimes.We

2
| METHODOLOGY AND MATERIAL 2.1 | Satellite imagery In order to assess the dune development within the study area, satellite images from different time periods were utilized.With freely F I G U R E 1 Overview map of the GB and the barchan dunes, mapped in this study.[Color figure can be viewed at wileyonlinelibrary.com] available data sources, like the Landsat 5, 7 and 8 and Sentinel-2 satellite imagery archives, a time scale as far back as 1984 can be covered (Earth Resources Observation and Science [EROS] Center, 2017b; Landsat Data AccessjU.S. Geological Survey, n.d.).These images have a spatial resolution of 30 m or less, making details of micro-scale landforms hard to identify.For the most recent time, ESRI basemap data were used.In order to extent the time span further into the past, we acquired CORONA satellite espionage images, specifically from the KH-4B (KH = keyhole) mission, which ran from 1967 till 1972 with a good coverage of northern China.The spatial resolution is up to about 1.8 m (Figure 2) according to the US Geological Survey (Earth Resources Observation and Science [EROS] Center, 2017a).The images were taken by two analog cameras, one forward-facing and one afterward-facing, resulting in a stereoscopic image setup (Earth Resources Observation and Science [EROS] Center, 2017a).The images were declassified and made available through the USGS in 1995.This archive holds great value for global land surface change research.Multiple studies have successfully used CORONA satellite images for different purposes . Especially at the edges of the image strips, the distortion can become quite severe.Therefore, as far as possible, only the sections towards the centre of the scanned film strips were used.After a rough corrections of rotation, scale and shift of the images, a detailed georeferencing was carried out.The whole georeferencing resulted in three mosaics from 25 KH-4B images, covering almost the entire GB.An average of almost 35 control points on the ESRI highresolution basemap was required to sufficiently georeference the images, applying a spline transformation.This transformation method yielded better results in accuracy then the second or third polynomial transformation, when checked against fix points identifiable on all images.The ESRI basemap is provided by Maxar, being a mosaic of WorldView and GeoEye satellite images from 2017 to 2020 in our area of interest (ESRI 2022).Moreover, as of early 2022, Landsat 9 images are available and were used for further validation of the mapped crestlines and to map anthropogenic structures, such as the photovoltaic power plant of the GB.In order to fill the time gaps between the 1968 CORONA images and the present ESRI basemap data, two Landsat TM scenes, from 15 August 1987 and 12 July 2001, were chosen, featuring a spatial resolution of 30 m in the bands covering the visible light.These Landsat images also dictate our time periods: 1968-1987 (T1), 1987-2001 (T2) and 2001-present (T3).
CORONA and the other images used (age, greyscale and spatial resolution), third party image detection, as applied in other studies F I G U R E 2 Example for the high spatial resolution of the 1968 CORONA KH-4B satellite images.[Color figure can be viewed at wileyonlinelibrary.com] (Baird et al., 2019; Havivi et al., 2018), was not feasible.Horizontal uncertainties of migration lengths were calculated for each time interval based on the varying ground resolution of the different satellite imagery: T1: ±0.15 m a À1 * 19 a = 2.85 m (1968-1987), T2: ±3.38 m a À1 * 14 a = 47.32 m (1987-2001) and T3: ±0.37 m a À1 * 19 a = 7.03 m (2001-2019).For the whole time span of 1968-2019, the uncertainty value is ±0.04 m a À1 * 52 a = 2.08 m.

2. 4 |
ERA-5 climate data and NDVI extractionERA-5 provides hourly and daily estimates on a number of climatic parameters, for instance, wind, precipitation and snow cover data(ECMWFjParameter database, 2023).All data are presented in a 30 km spatial grid and obtainable in either subdaily or monthly averages.The dataset is based on past observations, and modelling is used to reanalyse data as far back as 1950.The period of 1979 onwards is completed and updated on a regular basis (ECMWF Confluence Wiki, 2023).For the analysis, the dataset of daily aggregates was used (C3S 2017).From this dataset, the 10 m u-and v-component of wind, the 2 m mean air temperature and the total precipitation were spatially aggregated and plotted.Other climate model datasets have been compared in previous studies to station data as well as to the ERA-5 model, especially for precipitation and temperature(Arshad DP quantifies the overall potential of sand transport by wind.The resultant drift potential (RDP) represents the magnitude of the net resultant sand drift vector, and the resultant drift direction (RDDir) is its direction.The directional variability of the sand drift regime is quantified by the ratio of RDP/DP and is associated with dune shape types.High RDP/DP values indicates a unidirectional sand drift regime, promoting the formation of barchan and transverse dunes, whereas low values indicate a multidirectional regime, which is 04 to 23.8 ± 0.04 m a À1 were obtained.For the time intervals between, using the Landsat 5 and 7 images from 1987 and 2001, migration rates of 9.5 ± 0.15 m a À1 (1968-1987), 8.86 ± 3.38 m a À1 (1987-2001) and 7.69 ± 0.37 m a À1 (2001-present) were calculated, showing a decrease in dune migration velocity.The total distances travelled (1968-present) ranged from 26 m within the dune field on the western shore of the Longyangxia reservoir to 1238 ± 2 m for a singular dune in the north-western part of the basin.A decrease of migration rates from north-west to south-east of the GB is visible NDVI over the GB is based on Copernicus Sentinel-2 imagery of August 2021 was calculated.Highest values occur south-east of the Longyangxia reservoir, as well as towards the edges of the basin.The dunes and sand fields, as well as the water bodies, are well reflected by low NDVI values.A multi-temporal analysis of the yearly averages shows an increase of 0.16 to 0.19 from 1987 to 2019, respectively.The NDVI rises in all seasons (spring: March-May, summer: June-August, autumn: September-November and winter: December-February) in the two time periods 1987-2001 and 2001-2019.For the three clusters, NDVI analysis corridors were defined in a north-western direction, following the resultant drift direction (RDD) of the DP in order to assess the sand source from the prevailing winds (Figure 5).For cluster C3, two areas were taken into account to cover the whole eastern shore of the reservoir.Again, all seasons in both time slices in all four areas show a rise in the NDVI values (Table Figures S1, S2 and S3).The influence from the Westerlies in spring and winter results in stronger DP.In summer and autumn, the EASM winds coming from the south-east are weakening the RDP while not reversing the generally north-western wind direction.The RDP values are coming almost to a hold, however.This pattern is also represented within the RDDir values in the rose diagrams.
our study area over the past five decades are the barchans in the north-western part (16.65, 18.48 and 23.81 m a À1 for the past five F I G U R E 6 Precipitation and temperature in the GB from 1979 to present as recorded in the ERA-5 model data.The NDVI, starting from 1986, is shown below, calculated in GEE from Landsat 5, 7 and 8 images.Lastly, the dune migration rates for all three time periods, showing a slowing trend, are displayed above for comparison.[Color figure can be viewed at wileyonlinelibrary.com]T A B L E 3 Overview of DP, RDP, RDP/DP and RDDir values for the defined regions and time spans. decades).These dunes are almost completely solitary.The dunes are located in a narrow part of the basin, suggesting that they are affected by a channelling effect.Singular barchans have been described as unstable and are expected to either grow or shrink in height (Durán et al., 2011).The barchan dunes in Cluster C1 of the GB lost a considerable amount of height.The height reduction could indicate future shrinkage.The three clusters of dunes show different behaviours, when compared to the overall averages of dune migration.They clearly follow the observed north-west to south-east decrease in migration rates and RDP values (T1: Figure 9, T2 and T3: Figures S5 and S6) In area C2, the average migration rate is 7.89 ± 0.04 m a À1 , just above the overall average of 7.33 ± 0.04 m a À1 .The slowest migration rates are found in the contiguous dune fields bordering the western shore of the reservoir (C3).Here, the dunes are moving much slower with 2.2 ± 0.04 m a À1 as the average compared to a mean of 8.3 ± 0.04 m a À1

F
I G U R E 7 Scatter plot of migration rates and their respective RDP values for all three time periods.A trend is visible, as well as the regional distribution in the three classes, as described in the main text.[Color figure can be viewed at wileyonlinelibrary.com]F I G U R E 8 Rose diagram for cluster C1 in Period T3; calculated from spring values.RDDir represented by the red line, the green triangles represent directional magnitudes of DP and RDP.[Color figure can be viewed at wileyonlinelibrary.com] in order to detect the wind patterns from the EASM and the Westerlies.The dominating winds from the Westerlies are well represented in higher DP and RDP values during winter and autumn, as well as the ones from the EASM, with lower values in spring and summer.Also, we find a gradient of DP values decreasing from the north-western to the south-eastern areas, fitting to the dune migration rates.Further, when averaging the migration rates to the same clusters as analysed in the DP averages, we find a strong agreement.This shows a good representation of the mapped dune migration by the climate model data from ERA-5.However, the increasing rates we find in areas C1 and C3, are not reflected accordingly in the RDP averages.This indicates a high spatial variability across the whole basin as well as limited detectability of micro-to meso-scale patterns of climatic changes in the coarser resolution of the ERA-5 grid cells.On the other hand, the rates for the second time period are mapped only from Landsat images with a lower spatial resolution, potentially introducing a higher uncertainty, especially in area C3, where only low migration rates are found anyway.For all time periods and locations of our study area, we have observed RDP/DP values between 0.92 and 0.98, indicative for an almost unidirectional

F
I G U R E 9 RDP raster blended over the GB for period T1.The high values in the north are representing the Qinghai Lake, where high winds are common.The decrease from the north-west towards the south-east is well represented.[Color figure can be viewed at wileyonlinelibrary.com] conclude that the ERA-5 wind reanalysis data are well represented by the dune migration on the north-eastern TP.Using the KH-4B satellite images from 1968, enabled us to obtain migration rates of active barchan dunes.The mapped dunes in the GB are experiencing an overall decrease in migration rates by À1.81 m a À1 within the study period of the past 52 years.Similar reductions in migration speed are also found within every studied time interval.This is in accordance with the decreasing RDP values calculated from the ERA-5 climate model.Also, the rising NDVI values are indicating more sand fixations through vegetation.They are well correlated with the likewise rising temperature and precipitation, shown in the ERA-5 data.Overall, the dune migration rates and the ERA-5 climate model reflect the same development.This suggests a good agreement between the development of aeolian landforms and the ERA-5 climate reanalysis model also in high-altitude regions with complex topography.