Recent soil survey data, such as the LUCAS topsoil and RS data, have made large-scale soil erosion modeling more feasible. LUCAS Soil, which is an extensive and recurring topsoil survey conducted every three years across the European Union, is an example of such data [30]. The development of dynamic indexes and proxies for soil coverage will be facilitated by the increased availability of surveyed data and the utilization of RS data (MODIS, Landsat) and Copernicus products [31]. Landsat, with its collection of images dating back to the Landsat Multispectral Scanner (MSS) to the Landsat Operational Land Imager (OLI) and Thermal Infrared Sensor (TIRS) is home to the oldest archive of imagery from various sensors. With a spatial resolution of 30 m, it continues to be one of the most utilized satellite images in WE modeling, as shown in Figure 3. Additionally, the Shuttle Radar Topography Mission (SRTM) data, derived from synthetic aperture radar (SAR) imagery, remain a critical source of data for erosion assessment. Other sensors, such as ASTER, Sentinel, and SPOT, among others, offer lower or medium spatial resolution.

As illustrated in Figure 3, the number of indicators and RS sensors used through the WE evaluation is very variable. Different usage of RS data is based on its cost. Acquiring most of these images comes with a low cost or is entirely free in some cases. Additionally, high spatial resolution sensors such as IKONOS, GeoEye, and QuickBird can be utilized to evaluate erosion factors. Nonetheless, their usage in erosion assessment over large areas is often limited due to the high costs associated with acquiring them [32]. The MODIS data series was the most applied image. Among these studies, the utilization of freely available Landsat images is frequently observed, followed by images from Google Earth.

Erodibility, a crucial factor for predicting WE [33,34,35], refers to a soil’s susceptibility to erosion under specific meteorological conditions, or the efficiency of soil erosion on a surface given certain meteorological forcing. The interaction of fine soil particles (silt, clay, and sand) and organic carbon typically determines erodibility and is associated with factors such as soil structure, organic content, surface roughness, and soil texture [36]. However, measuring erodibility can be challenging due to these multiple factors. The influence of soil surface conditions based on soil type is often overlooked as a source of variation in characterizing soil surface erodibility and in using RS for soil assessment [37]. The computation of soil erodibility has been approached differently in various studies [38]. Ref. [39] calculated the wind erodibility of European soils by using a multiple regression equation based on soil texture and chemical properties, while Ref. [40] used soil organic carbon and soil particle size distribution to calculate erodibility in Inner Mongolia, China. However, soil erodibility equations typically require specific physical soil properties that may not be widely available. In the United States, the wind erodibility index (WEI) is available through the USDA Web Soil Survey geographic database [41]. Conducting long-term erosion plot studies is necessary to obtain direct measurements of soil erodibility. Most of the earlier studies [42,43,44,45] have been limited only to experimental study areas. Ref. [46] analyzed the soil erodibility of soil moisture, gravel cover, and surface roughness in combination with Landsat8 data to evaluate the direct impact on soil erodibility. Close-range photogrammetry was used to generate a gravel percentage map and the ASTER digital topographic model was used to analyze surface roughness. Ref. [47] highlighted the potential of directional and multi-angular soil reflectance to enhance the knowledge of erodibility and to detect and quantify soil erosion. In addition, the authors suggested that utilizing ground-based radiometers, as well as current and upcoming versions of angular sensors mounted on aerial and satellite platforms, and multi-angular spectral reflectance holds promise for remotely assessing and computing soil surface characteristics, thereby offering a comprehensive framework. Bi-directional soil spectral reflectance models [48,49] have the potential to establish a framework for integrating the analysis of soil surface erodibility and potentially measuring WE. Ref. [50] employed Landsat MSS and AVIRIS imagery to illustrate the movement of sand from deserted agricultural areas.

Remote sensing is a valuable and economical means of obtaining data for mapping and modeling soil properties, soil resources, and their temporal and spatial variations [47,51,52,53,54]. Utilizing Visible, Near Infrared, and Shortwave Infrared sensors, it is possible to calibrate site-specific datasets that depict the correlation between spatiotemporal and quantitative variations in soil information across extensive regions [53]. Ref. [55] estimated soil erodibility by the support vector regression method on Landsat-8 images with relative percent deviation equals 2.01.

The presence of vegetation, crop residues, or clouds can hinder the accuracy of soil property predictions, which is why extracting spectral information solely from bare soil pixels may yield successful results [56]. Ref. [57] stated that a single satellite image provides only 0.5% of valid bare soil pixels, which is inadequate for soil mapping. To address this, new methods of bare soil detection, especially from Landsat and Sentinel images, are being employed. Ref. [58] used the MID-Infrared profile of four images to identify bare soil pixels and established a correlation between the amount of clay content and these pixels to generate a soil clay content map. Refs. [59,60] used a fusion technique, which increased the exposed soil coverage from 36% (single image) to 75%. Meanwhile, Ref. [61] utilized a five-year compilation of Landsat TM imagery and the Barest Soil Composite Image (BSCI) technique to achieve 90% in soil mapping. In a similar attempt, Ref. [62] integrated the reflectance and spectral parameters obtained from Sentinel-2 images with a wider range of spectral and non-spectral covariates to forecast topsoil clay content. Ref. [63] also proved that multi-temporal images performed better in predicting soil properties, then single-date images. The Soil Composite Mapping Processor (SCMaP) was devised by [64] to surmount the obstacle of insufficient soil exposure and to map soil properties in Germany. Further research advancement and improvement is expected as high signal-to-noise ratio spectrometers on space-borne imagery become available, along with the PRISMA, such as EnMAP [65], HISUI [66], SHALOM [67], HySpex [68], and the Sentinel-10/CHIME [69].

The presence of soil moisture is a significant contributor to the cohesion of soil particles and acts as a mitigating factor for wind erosion [16]. Soil erodibility reaches its peak when the soil is completely dehydrated. However, as the soil moisture content increases beyond this limit, there is a noticeable reduction in erodibility. Erodibility rapidly increases with the soil moisture content [70]. Precisely monitoring and estimating the spatial and temporal changes in soil moisture is of great significance [71]. Estimating soil moisture becomes more challenging in regions with dense vegetation or snow cover, as well as areas with intricate topography [72]. Soil moisture content in correlation with the number of erosive days could predict some mitigation effect for some areas [39].

The production of dust is connected to the annual diversity in plant cover resulting from the variation in soil moisture [50]. Traditional techniques of restoration and reclamation necessitate surface disturbance and rely on adequate moisture to avoid erosion, making them vulnerable to significant erosion risk in the absence of wet conditions [16]. Several machine learning algorithms have been utilized for extracting water bodies and shorelines from remote sensing images, including neural networks, convolutional neural networks, artificial neural networks, support vector machines, naive Bayes, random forest, gradient boosted machine, recursive partitioning and regression trees, and constraint energy minimizations [73]. Individual spot measurements do not capture high spatial and temporal variability [74,75] and therefore are not suitable for regional and global research [71,76].

Surface roughness refers to the way surfaces are expressed in terms of topography, including features on horizontal and vertical scales ranging from millimeters to a few hundred meters [29]. Precise understanding of surface roughness and its influence on wind speed is crucial for various fields, including climate modeling, wind power meteorology, agriculture, and erosion hazard assessment [89]. Nearly flat surfaces with loose sand-sized particles, lacking any shelter or cover objects, are particularly sensitive to WE [90].

Surface roughness can be considered a direct input parameter or indirectly estimated through soil moisture or vegetation biomass in WE modeling [91,92]. It can be utilized as a WE predictor by determining the potential for soil particle retention, emission, and saltation [93]. However, the measurement accuracy of surface roughness is constrained by the measurement method, instruments used, and data pre-processing [94]. Surface roughness can be measured using contact or non-contact methods. Contact methods include the grid plate and profiler methods, while non-contact methods include laser profiler, stereophotography, pencil-beam radiometer, and LiDAR [95,96]. Achieving the required measurement accuracy is challenging due to limitations such as large spatial sampling intervals, short sampling segments, temporal variability, and inconvenient operation. However, TLS as a non-contact device [97] can furnish spatial data of high resolution (in mm), depicting surface alterations over time [98]. Other surface properties can be derived from the strength of the reflected signal. The best machine learning approaches for surface roughness estimation consist of both classical machine learning algorithms such as random forest, neural network and support vector machine and artificial neural network (ANN), as well as deep learning algorithms including multilayer perceptron (MLP) and deep belief network (DBN) [99].

Vegetation reduces WE as follows: 1. safeguards the soil from erosive influences by enveloping a portion of its surface; 2. diminishes the velocity of the wind near the ground; 3. traps the soil particles [107]. Vegetation represents mainly standing live and dead biomass (e.g., crops in various development stages) and standing or flat litter (e.g., crop residues). The impact of vegetation on soil erodibility is contingent upon various factors such as the extent of vegetation cover, height, and spatial distribution or arrangement within a landscape [108]. The establishment of vegetation cover and development of windbreaks are the most common erosion control measures [18].

The impact of vegetation on WE in drylands has been primarily studied through wind tunnel experiments, windbreak studies, and field trials [107]. The degree of vegetation’s influence on WE is determined by the amount, height, and distribution of vegetation in the landscape. During different seasons, the state of vegetation cover undergoes changes, which in turn affects its ability to mitigate wind erosion. The impact of seasonality on vegetation cover relates to the growth and development stages of crops or plant species. In some regions, vegetation cover may decrease during dry seasons or winter, leading to reduced erosion control capacity. Conversely, during the rainy seasons or periods of active growth, the vegetation cover tends to increase, providing greater protection against erosion. The extent of vegetation cover is often represented as the ratio of the surface area covered by plants [109]. If the vegetation cover is below 20%, it has a minimal impact on reducing wind velocity, while erosion is essentially halted if the vegetation cover exceeds 60% [110]. Therefore, understanding the seasonality of vegetation cover is crucial for assessing its effectiveness in erosion control throughout the year. The review focusing on erosion monitoring [111] identified Landsat, SPOT, and MODIS data as the most-used RS systems for state-based monitoring of ground cover and identified the products most suitable for erosion monitoring [112,113,114]. For the WE modelling, the information about vegetation cover, its annual dynamics, and crop residues are important. Vegetation cover for the European scale model [115] was estimated from the Leaf Area Index derived from ENVISAT/MERIS satellite images. For the European WE model (WEELS), the Land Use data were gathered from existing maps, farmers files, agricultural statistics, or the land use dataset was estimated through modeling techniques in case of unavailability [36]. In [116]’s study on the East Africa region, the vegetation cover fraction was obtained from GCLS SPOT/PROBA-V images at a resolution of 1 km. For the East Africa region, Ref. [116] used a fraction of vegetation cover derived from GCLS SPOT/PROBA-V images at 1 km resolution. For the regional studies, the data from MODIS with a resolution of 250 m or 1 km [117,118,119] or Landsat with a resolution of 30 m [120,121] are often used. The pre-processed data products from Copernicus Global Land Services are also used [122]. For the historical analyses of the vegetation cover development, the MODIS data were combined with 8 km spatial resolution AVHRR [40,123,124] or the Landsat data were used [125].

Vegetation cover is often derived from the Normalised Difference Vegetation Index [126] using a dimidiate pixel model [127] or a specific regression model is developed [122]. From the range of vegetation indexes the Average Maximum Modified Soil Adjusted Vegetation Index was also used [120,128]. A classical RS classification approach was also used for the vegetation cover estimation [125]. In small experimental areas, the vegetation cover could be visually mapped in the field [129], using the line transect method [130] or a pin-type profile meter [131]. The impact of vegetation cover can be disregarded when examining the potential for WE [132], or in the case where the erosion events are expected at a time when there is no vegetation cover in the field [133]. Support Vector Regression and Random Forests are two popular classic machine learning algorithms, and have been widely used in regression and classification of vegetation-related studies [134].

The windbreaks have various functions in landscapes. Beside the soil protection against wind and water erosion, they positively affect the microclimate, plant productivity [150,151], soil properties [152,153,154], biodiversity, and also serve other purposes such as protecting buildings, offering thermal comfort to livestock, reducing snow and sand accumulation, and mitigating wind impacts on the transportation infrastructure [155]. Therefore, the effect of windbreaks is evaluated from different perspectives [156,157]. The WE model incorporates three aspects of wind barriers: (1) wind barriers’ height; (2) optical porosity; (3) wind barriers’ width. The wind barriers’ height affects the shelter length [158]. For the erosion modelling purpose, the height is measured by altimeter [159] or from vertical images of the windbreak [132], or is estimated from the average height of windbreak types [160]. Ref. [161] estimated the average tree height of windbreaks from first principal component of GF-1 Multi-Spectral high resolution satellite image. Porosity refers to the proportion of open space to the occupied space of tree stems, branches, twigs, and leaves. It impacts the degree of wind speed reduction and shelter effect of windbreaks [162]. Since measuring aerodynamic porosity is challenging, optical porosity is used as an alternative. Optical porosity is calculated as the ratio of white pixels to the total pixels in digitized black and white pictures taken at a distance of 30 m from the windbreak [163]. These images are processed using software such as ImageTool 3.0 or GIMP 2.10.34 [132,159,164,165]. In RS-based windbreak studies, it is common to assume an approximate shape for the simulated trees. However, this approach often results in notable discrepancies between the simulated and actual results. Ref. [166] employed TLS to obtain parameters used in a quantitative structural model (AdQSM) to reconstruct the tree structure and restore the wind field environment using the computational fluid dynamics software PHOENICS-2022. The findings suggest that TLS can effectively retrieve windbreak parameters and quantify porosity, thereby substantially improving the credibility of windbreak impact investigations in windbreak forests. The wind barriers’ width is related to barrier porosity [161,165], but also has a specific effect on wind velocity [167]. Using 10 m SPOT5 images is applicable for windbreaks’ width measurement; however, errors may arise when the windbreak is presented on both sides of the road, in the case of sparse or young windbreaks [168]. Aerial images were analyzed using object-oriented image analysis to map the windbreaks [169,170,171], but also for the automatic extraction of wind barriers’ width [172]. The wind barriers are neglected in regional scale models [17,35,39,116,122,173], because such a dataset does not exist for large areas. In most of the local studies, the effect of wind barriers is modelled in the direction of prevailing winds [159,165] or as a Euclidean distance from the nearest windbreak [161]. The most innovative methodology is to calculate the wind reduction for all wind directions and to weight them according to the average occurrence of erosive winds [132]. A unique method is to estimate the shelter area from lateral cover and shadows derived from the multitemporal MODIS albedo images [147].

RS approach is used more often for deriving the inputs for WE models, than for mapping the eroded areas. The sole effect of WE is hard to recognize from RS images; therefore, the compound effect with water erosion [187], tillage erosion [188], or land degradation in general [20,21] is mapped.

Monitoring of agricultural soil degradation by RS was reviewed by the authors of [21] and they identified two RS approaches for mapping the WE: (1) mapping direct indicators such as surface lowering or changes in surface roughness; and (2) mapping indirect indicators based on WE relationship with changes in vegetation cover. One illustration of the initial approach involves utilizing radar remote sensing to delineate the patterns of wind-induced land deterioration in the northeast Patagonian region [189]. A second approach could be illustrated by a study that used ASTER images to obtain a WE risk map for Inner Mongolia from the basis of the land use classes, and vegetation height [190].

Ref. [187] summarized the contribution of RS to mapping land depletion in Latin America and Caribbean. The most-used WE indicators were sand dunes, however various surface erosion features such as wind streaks, desert pavements, paleo–aeolian sand features, sand encroachments, blowouts, and changes in vegetation cover were used as well.

Machine learning models commonly employed for wind erosion mapping in various environmental research areas encompass decision tree models, linear equation models, particle swarm optimization-adaptive network-based fuzzy inference system, genetic algorithms, support vector regression, artificial neural networks, hybrid models, random forest, Wang and Mendel’s method, partial least square regression, principal component regression, Cubist, Bayesian additive regression trees, radial basis function, extreme gradient boosting, and regression tree analysis [188].