Introduction

Increasing the share of renewable energy is essential to realize the global emission reduction targets1,2. According to the current global emission reduction trends, it is difficult to achieve the global 1.5 °C/2 °C temperature increase goals and 2050/2070 net-zero emission targets3,4,5. To reduce anthropogenic carbon dioxide (CO2) emissions, the exploration of renewable energy at a global scale need be strengthened1,6,7. In recent years, solar energy – as affordable and clean energy – has been increasingly utilized8. A large number of solar farms have been built across the globe8,9. Deserts with low land value and long sunshine time are favorable for building solar farms10,11. In turn, solar farms in deserts can increase surface friction, reduce surface albedo, enhance local precipitation, and increase regional vegetation in and around deserts10. Hence, desert solar geoengineering should be considered a feasible action program of planetary geoengineering12,13 aiming at mitigating anthropogenic greenhouse gas emissions. For building desert solar farms, the existing site suitability methodologies14,15,16 cannot effectively solve the dune threats (e.g. sand burial and dust contamination) to solar photovoltaic panels across global deserts.

Dune threats are associated with sand flux, and sand flux driven by effective shear velocities reflects the potential sediment transport capacity of the wind17,18,19,20,21,22,23,24. Sand flux in this study can be briefly quantified through the flux potential (FP) and resultant flux potential (RFP). This is similar to the drift potential and resultant drift potential of sand drift25,26,27, the absolute potential sand flux and resultant potential sand flux18,19,20. FP is the sum of bulk fluxes in all azimuths, and RFP is calculated by the Euclidean formula of the projected due-north and due-east bulk flux components from all azimuths28 (METHODS). Note that the flux calculation is for the saturated flux. The true flux may be smaller (due to precipitation or erodible surface fraction) or larger (due to dune steepness), but this is a reasonable estimate with precedents in other studies18,19,20,21. FP and RFP of sand flux have been used to quantify dune activities18,19,20,21. Theoretically, FP represents wind energy, so higher FP means greater transport capacity of instantaneous winds in all azimuths; RFP represents the net sand transport potential in the resultant flux direction, so higher RFP means severe accumulation25,26; FP is more important than RFP in assessing the dune threats. Most studies of sand flux are based on the wind data from local meteorological stations29. However, global meteorological stations are limited in many desert areas27. Wind data from the reanalysis products with different spatiotemporal resolutions provide a feasible scheme for quantifying sand flux at a global scale18,19,20,21. For example, the ERA5 reanalysis product (0.25° × 0.25° resolution)30 was used to calculate the FP and RFP of sand flux18,19,20. Accordingly, the one-hour-scale instantaneous wind data from the ERA5-Land reanalysis product with a higher resolution (0.1° × 0.1°)31 should be able to adequately capture more spatial details of sand flux changes21, and then assess the dune threats to desert solar farms. However, how to use the FP and RFP to effectively optimize the site selection of solar farms across global deserts remains unsolved.

In this study, we resample desertified lands and sandy lands at 500 m resolution (extracted by the support vector machine analysis, trial-and-error method and visual interpretation analyses based on the Moderate Resolution Imaging Spectroradiometer data)32 into global deserts at 0.1° × 0.1° resolution (Fig. 1). We use the eastward and northward wind components at the height of 10 m from the ERA5-Land hourly wind data to calculate the yearly sand flux for the period 1950–2022, and adopt the 73-yr mean sand flux to assess the suitability of global deserts for building solar farms. According to solar farm scores, we can reduce or avoid the dune threats, and efficiently operate desert solar farms.

Fig. 1: Spatial distribution of global deserts.
figure 1

Deserts are resampled to a resolution of 0.1° × 0.1°, matching the spatial resolution of the ERA5-Land hourly wind data. The colored abbreviations are the three-letter ISO 3166–1 alpha-3 GADM country codes. The countries in Asia are colored by the malachite green, the countries in Africa the mars red, the countries in South America the ginger pink, the countries in North America the moorea blue, the countries in Europe the cretan blue, and the countries in Australasia the anemone violet. The boundaries of the country with desert data are colored by 50% gray, and the rest are colored by 10% gray.

Full size image

Results

73-yr mean sand flux

The data representing global deserts with 0.1° × 0.1° resolution were distributed in 55 countries, including 23 countries in Asia, 20 countries in Africa, 4 countries in South America, 2 countries in North America, 1 country in Europe and 1 country in Australasia (Fig. 1).

We calculated the yearly FP and RFP from the ERA5-Land hourly wind data (METHODS). During 1950–2022, the FP mean of global deserts was 23.7 ± 3.9 m3 m-1 yr-1 (mean ± standard deviation), with the maximum mean and standard deviation of 282.1 m3 m-1 yr-1 and 26.5 m3 m-1 yr-1 on the ERA5-Land grid-scale, respectively. The FP means had patchy distribution globally. In terms of the ERA5-Land grid point number, the FP means of 0–20 m3 m-1 yr-1 were dominant, and followed by the patches of 20–40 m3 m-1 yr-1. The FP means greater than 40 m3 m-1 yr-1 are shown in Fig. 2a.

Fig. 2: The (a) FP and (b) RFP means of global deserts for the period 1950–2022.
figure 2

The equidistant spacings of the FP and RFP means are set to 20 m3 m-1 yr-1 and 1 m3 m-1 yr-1, respectively.

Full size image

The RFP mean of global deserts was 0.7 ± 0.4 m3 m-1 yr-1, with the maximum mean and standard deviation of 11.8 m3 m-1 yr-1 and 4.1 m3 m-1 yr-1 on the grid-scale, respectively. The RFP means also had patchy distribution across global deserts. Most deserts were dominated by the patches with the RFP means of 0–1.0 m3 m-1 yr-1, and then the patches of 1.0–2.0 m3 m-1 yr-1. The patches with the RFP mean greater than 2.0 m3 m-1 y-1 are shown in Fig. 2b. The patches with high RFP mean may have high dune celerities28,33. The spatial distributions of the FP and RFP standard deviations can be seen in Supplementary Figure 1.

In this study, the spatial distributions of the 73-yr mean FP and RFP calculated by the one-hour-scale instantaneous wind data from the ERA5-Land reanalysis product were similar to those of the 15-yr mean drift potential and resultant drift potential calculated by the fifteen-minute-scale instantaneous wind simulations from the HadGEM3-GC3.1 model family for the period 2000–201527. This suggests that the interpolation from ERA5 to ERA5-Land hourly wind data31 does not filter out high wind speed events27, and the ERA5-Land hourly wind data effectively capture the basic characteristics of sand flux across global deserts.

Scoring scheme for desert solar farms

We classified the 73-yr mean sand flux to construct a scoring scheme. First, the FP and RFP means were used to quantify the sand burial degree, and the FP means were used to distinguish the dust contamination degree. Then, we divided the FP means and the RFP means into 4 classes separately using quartile classification (Fig. 3a,b), intersected the FP mean classes and the RFP mean classes, removed the non-observed permutations and scored the suitability according to the applied rule, in which we assumed that the FP mean is more important than the RFP mean in scoring the suitability of global deserts due to low solar photovoltaic panels (METHODS).

Fig. 3: The scoring scheme and result for solar farms based on changes in sand flux.
figure 3

First intersect the (a) FP and (b) RFP mean classes, then remove non-observed permutations, and finally apply the simple rule to assign the corresponding scores for solar farms (c). The left insets show the percentage distribution of geodesic area for solar farm scores.

Full size image

The first step of the scoring scheme is to divide the FP means into 4 classes using the FP mean quartiles: the first quartile (13.2 m3 m-1 yr-1), the median (21.2 m3 m-1 yr-1) and the third quartile (30.1 m3 m-1 yr-1). These classes had the geodesic area of 2706.9 × 103 km2, 2720.4 × 103 km2, 2665.0 × 103 km2 and 2638.8 × 103 km2, respectively (Fig. 3a). The second step is to divide the RFP means into 4 classes using the RFP mean quartiles: the first quartile (0.5 m3 m-1 yr-1), the median (0.6 m3 m-1 yr-1) and the third quartile (0.8 m3 m-1 yr-1). These classes had the geodesic area of 2736.8 × 103 km2, 2730.0 × 103 km2, 2649.9 × 103 km2 and 2614.2 × 103 km2, respectively (Fig. 3b). The final step is to intersect the FP and RFP mean classes. We removed the non-observed permutations and got the scores of solar farms according to the applied rule. The ascending FP and RFP mean classes are unfavorable for solar farms (Table 1, more details see METHODS).

Table 1 Solar farm scores across global deserts
Full size table

Solar farm scores based on quartile classification of the FP and RFP means showed obvious patchy distribution across global deserts. For solar farms, the highest score 15 had the maximum grid point number of 21068 and geodesic area of 2333.3 × 103 km2. In contrast, score 12 had the minimum grid point number of 1 and geodesic area of 0.1 × 103 km2. For the rest, see Fig. 3c inset and Table 1. If only consider the dune threats, high (low) scores clearly showed that global deserts had strong (weak) suitability for building solar farms. In conclusion, the criteria of site selection for solar farms varied across the globe.

Discussion

Our results demonstrate heterogeneous spatial distribution of sand flux and wind environment classifications of global deserts, and present a scoring scheme for the site selection of solar farms across global deserts on the basis of the 73-yr mean sand flux that reflects the basic characteristics of sand flux. In this study, we assumed that the FP mean is more important than the RFP mean in evaluating the threats to low solar photovoltaic panels. FP is intercepted by solar photovoltaic panels because a solar farm represents a local sink area. High FP brings severe sandblasting34,35 and causes severe dust contamination on solar photovoltaic panels. RFP causes the sand burial of solar photovoltaic panels in the resultant flux direction. In addition, we adopt the quartile classification of the FP and RFP mean distributions to ensure the logical rationality of the scoring scheme. Furthermore, we find 47.2% of the existing solar installation sites36 in deserts are located in the highest-score regions of solar farms (Fig. 4 and Table 1). The inconsistency of score orders sorted by area percentage and scoring frequency also reflect the robustness of our scoring scheme (Figs. 3, 4 and Table 1).

Fig. 4: Validation of solar farm scores in deserts.
figure 4

The black solid squares represent the locations of the 216 solar installations in deserts. The inset presents the scoring frequencies extracted by the existing solar installations in deserts.

Full size image

This study provides a guide to select the regions suitable for desert solar farms. Using the wind data from the reanalysis products with different spatiotemporal resolutions18,19,20,21, especially, the ERA5-Land reanalysis product (0.1° × 0.1° and hourly resolution)21, could detailedly characterize the wind environments and quantify the dune threats at a global scale. In this study, we neglect the errors introduced by the interpolation from the ERA5 to ERA5-Land hourly wind data, especially in complex terrains or coastal areas31. Some deserts have no effective shear velocities and small or zero flux18,19,20,26. They may be interpreted as the ancient dune systems or be driven by other episodic factors (e.g. alluvial/fluvial, lacustrine and coastal). But this study only focuses on the potential sediment transport capacity determined by effective shear velocities17,18,19,20,21,22,23,24. In the actual site selection, local situations such as sediment availability37, topographic influences38,39 and precipitation effect19,40 should also be considered.

Our scoring scheme could be used to choose the best sites for solar farms in the regions affected by dune threats, and also to assess the site selection of traffic engineering, petroleum exploitation and irrigated farming in desert environments. Furthermore, our results can help improve desert solar geoengineering and achieve the Sustainable Development Goal 7 (“affordable, reliable, sustainable and modern energy for all”) by 203041, and may even indirectly contributes to maintaining the global surface temperature increase below 1.5–2 °C and reaching the global carbon neutrality42 over the long-term.

Methods

Desert data

Wu et al., 202232 extracted the Moderate Resolution Imaging Spectroradiometer (MODIS) Terra MOD09A1 product with 500 m resolution from the bare ground areas during 201543. Next, they used the independent components analysis tool to enhance the spectra of the mosaiced MOD09A1 product. After that the support vector machine method trained on 80612 samples was employed to extract the desert areas, achieving a classification accuracy of 79.83% for 50226 test samples. Then, they used the trial-and-error method to extract the areas with the relief degree ≤500 m. This improved the classification accuracy to 81.87%. Later, they used the high-resolution satellite image of Google Earth to visually interpret and identify the areas that cannot be distinguished by machine learning. By doing this, the classification accuracy of desert areas reached to 92.37%. Finally, land cover types in desert areas included grassland, shrub, desertified land and sandy land, and gobi covers32. In this study, we referred to desertified lands and sandy lands as global deserts, and assumed that global deserts are covered by medium-to-fine sands.

We used the 73-yr mean FP as the snap raster to resample desertified lands and sandy lands with 500 m resolution into global deserts at 0.1° × 0.1° resolution, the same to the spatial resolution of the ERA5-Land reanalysis product from the European Center for Medium Range Weather Forecasts (ECMWF)31. The grid point number and geodesic area of global deserts were 98380 and 10731.0 × 103 km2, respectively.

Wind data

Wind data are from the eastward and northward wind components at the height of 10 m of the ERA5-Land reanalysis product, which has the hourly temporal resolution and 0.1° × 0.1° spatial resolution31. In this study, the ERA5-Land hourly wind data spanned from 1950 to 2022. The instantaneous wind speed \(U\) and azimuth \(A\) at the height of 10 m is calculated as

$$U=\root2\of{{{\bf{u}}}^{2}+{{\bf{v}}}^{2}}$$
(1)
$$A={atan}2({\bf{u}},{\bf{v}})$$
(2)

where \({\bf{u}}\) is the eastward component, in m s-1; and \({\bf{v}}\) is the northward component, in m s-1. Note that the ERA5-Land eastward and northward wind components are got by simply linear interpolating the ERA5 eastward and northward wind components based on a triangular mesh. They are not model output of the ECMWF land surface model at 0.1° × 0.1° resolution31.

Conceptual framework of sand flux

The shear velocity \({u}_{* }\) in m s-1 is calculated as

$${u}_{* }=\frac{U\kappa }{\mathrm{ln}\left(z/{z}_{0}\right)}$$
(3)

where \(U\) is the instantaneous wind speed at the height of 10 m, \(\kappa\) = 0.4 is Von Kármán constant, \(z\) = 10 m is the height above the Earth surface, \({z}_{0}\) = 0.001 m is the assumed roughness length above the sand surface44.

The impact threshold shear velocity \({u}_{* t}\) = 0.231 m s-1 is calculated by

$${u}_{* t}=\frac{\root{2}\of{{gd}{\rho }_{s}/{\rho }_{f}}}{10}$$
(4)

where \(g\) = 9.81 m2 s-1 is gravity acceleration; \(d\) = 0.00025 m is the reference median grain diameter of medium-to-fine sands for active deserts45,46,47; \({\rho }_{s}\) = 2650 kg m-3 is sand density; \({\rho }_{f}\) = 1.22 kg m-3 is air density23,45.

The saturated bulk flux \({{\bf{q}}}_{{\boldsymbol{b}}}\) in m3 m-1 s-1 is approximately22,23,24

$${{\bf{q}}}_{{\boldsymbol{b}}}=C\frac{{u}_{* t}}{g{\rho }_{b}}{\rho }_{f}\left({u}_{* e}^{2}-{u}_{* t}^{2}\right)$$
(5)

where \(C\) = 5 is an empirical (dimensionless) scaling parameter, \({\rho }_{b}\) = 1580 kg m-3 is the mean of bulk densities in other studies48,49,50,51,52,53,54,55,56,57,58,59 (Supplementary Table 1), \({u}_{* e}\) = \({u}_{* }\) > \({u}_{* t}\) is the effective shear velocity.

After deriving the effective shear velocity, and considering the intermittence of instantaneous winds, we also define \({{\bf{q}}}_{{\boldsymbol{b}}}\) = 0 when \({u}_{* }\) ≤ \({u}_{* t}\), and finally apply zero flux to the mean of the subsequent flux calculations, so the mean hourly flux vector lengths \({Q}_{b}\) in m3 m-1 s-1 and the mean hourly flux vectors \({{\bf{Q}}}_{{\boldsymbol{r}}}\) in m3 m-1 s-1 27,28,29 are given by

$${Q}_{b}=\frac{{\sum }_{i=1}^{N}\left|\vec{{q}_{b}}\right|}{N}$$
(6)
$${{\rm{Q}}}_{r}=\root{2}\of{{\left(\frac{{\sum }_{i=1}^{N}{{\bf{q}}}_{{\boldsymbol{b}}}\sin A}{N}\right)}^{2}+{\left(\frac{{\sum }_{i=1}^{N}{{\bf{q}}}_{{\boldsymbol{b}}}\cos A}{N}\right)}^{2}}$$
(7)

where \(N\) is the number of hours in a Julian year (8760 h for a common year and 8784 h for a leap year), and it represents the 8760 or 8784 instantaneous wind vector measurements19 with the one-hour sampling rate28. We only focus on the ERA5-Land hourly wind data in this study. This means we do not consider the influence from different temporal resolutions or different averaging time intervals of other wind data27,29. However, the one-hour-scale instantaneous wind data may underestimate the true bulk flux, because it cannot capture high wind speed events as effectively as the ten-minute-scale standard meteorological data27,29.

Finally, the flux potential (FP) and resultant flux potential (RFP) measured in m3 m-1 yr-1 are defined as

$${FP}=S{Q}_{b}$$
(8)
$$R{FP}=S{Q}_{r}$$
(9)

where \(S\) = 31536000 is the conversion factor from second to year (365 days). FP (a scalar value) is the sum of bulk fluxes in all azimuths, and it represents the transport capacity of instantaneous winds in all azimuths. RFP (a net resultant vector) is the Euclidean sum of the projected due-east and due-north bulk flux components from all azimuths, and it represents the net sand transport potential in the resultant flux direction, which is the net trend of sand flux, in line with the dominant direction of dune celerities28,33. We used the absolute RFP, neglecting its vector property.

In addition, the naming directions of FP and RFP follows where the sand moves. Eventually, RFP represents net sand transport potential under effective shear velocities in all azimuths on the ERA5-Land grid-scale after one full year60.

Calculating the 73-yr mean sand flux

We estimated the spatial distributions of the FP and RFP means across global dunes. Considering the uncertainty of wind speed from the ERA5-Land hourly wind data31, we extracted the spatial distributions of the standard deviations of the FP and RFP means during the study period (Supplementary Figure 1).

Area-weighted aggregated statistics

The means ± standard deviations of FP and RFP for global deserts were weighted by the grid cell area at a global scale, employing the CDO software61.

Rule of the scoring scheme

For interpretation and application, we divided the FP mean and the RFP mean into 4 classes separately using quartiles. The quartiles of the FP means were 13.2 m3 m-1 yr-1 (the first quartile), 21.2 m3 m-1 yr-1 (the median) and 30.1 m3 m-1 yr-1 (the third quartile). For the RFP means, the quartiles were 0.5 m3 m-1 yr-1 (the first quartile), 0.6 m3 m-1 yr-1 (the median) and 0.8 m3 m-1 yr-1 (the third quartile). For solar farms, FP reflects both the potential sand burial degree in all azimuths and the dust contamination degree on solar photovoltaic panels. High FP brings sandblasting34,35, and produces dusts that cover solar photovoltaic panel surface, reducing the solar photovoltaic conversion efficiency62. RFP reflects the potential sand burial degree of low solar photovoltaic panels in the resultant flux direction.

In our scoring scheme, due to low solar photovoltaic panels, we assigned greater importance to the FP mean over the RFP mean when scoring the suitability of global deserts. Higher FP and RFP means indicate less favorable conditions for solar farms. On the basis of the above empirical judgement, we applied one simple rule for scoring the suitability of geometric intersections between the FP and RFP mean classes.

We tabulated the solar farm score by the importance of empirical judgment about solar farms. The permutation number of the FP and RFP mean classes in sand flux was 16 (4 × 4).

The scoring scheme for solar farms included the following steps:

Step 1: Sort the FP mean class in the first column from high to low.

Step 2: In the second column, we still sequentially nested the RFP mean class from high to low under individual classes of the FP mean (from high to low).

Step 3: Considering the empirical judgment about solar farms, we assigned the score from 1 to 16. However, only one permutation was not observed at a global scale. We removed this permutation and reassigned the final solar farm score from 1 to 15 (Table 1).

Validation of the scoring scheme

The locations of solar installations used for the validation are from the global, open-access, harmonized spatial datasets based on the OpenStreetMap infrastructure data36,63. We used the desert data to mask the point vector data titled by the global_solar_2020, and identified the actual locations of solar installations in deserts (Fig. 4 and Table 1), in order to validate the robustness of our scoring scheme for solar farms in deserts.