From soil loss to sediment delivery: GeoAI-enhanced RUSLE modeling of erosion and sediment dynamics in a hyper-arid watershed

Abstract

In the delineated watershed containing Wadi Samnan near Az Zulfi, Saudi Arabia, soil erosion is a local management concern because sparse vegetation, erodible sandy surface materials, escarpment-influenced terrain, and episodic rainfall events can concentrate runoff and sediment movement along wadi channels and drainage corridors. These conditions make it difficult to identify erosion-prone zones using field observations alone, particularly in data-scarce hyper-arid environments. This study presents a GeoAI-enhanced RUSLE-based framework to model water-induced soil loss and sediment delivery dynamics in the study watershed. The Revised Universal Soil Loss Equation (RUSLE) was integrated with geospatial datasets, remote sensing products, digital elevation model-derived terrain attributes, and machine-learning-based spatial analysis to estimate the spatial distribution of water-induced erosion risk. Rainfall erosivity, soil erodibility, topographic influence, land-cover conditions, and conservation-practice factors were derived and mapped within a GIS environment. In addition, the Sediment Delivery Ratio (SDR) was incorporated to evaluate relative sediment-delivery potential and improve the interpretation of how estimated hillslope soil loss may be transferred into sediment yield within the drainage system. The results show that RUSLE-derived potential soil-loss estimates ranged from 0 to 319.5 t ha^-1 yr^-1, with the slight erosion class covering 43.2% of the watershed and modeled severe-erosion hotspots covering 10.9%. High-risk areas were mainly concentrated along steep slopes, drainage corridors, sparsely vegetated surfaces, and erodible sandy soils. SDR results indicated low overall sediment connectivity, with approximately 85%–99% of eroded material likely retained locally within the watershed, depending on the spatially varying SDR values. The Random Forest internal consistency analysis explained 81.73% of the variance in RUSLE-derived soil-loss estimates, indicating that the input factors captured much of the modeled spatial variability; however, this result should not be interpreted as independent field validation. Overall, the integration of RUSLE, SDR, remote sensing, GIS, and GeoAI provides a useful framework for identifying erosion-prone zones, distinguishing soil-loss potential from sediment-delivery potential, prioritizing conservation interventions, and supporting sustainable watershed management in hyper-arid environments.

Highlights

1 Introduction

Soil erosion and land degradation remain major environmental challenges in dryland regions, where fragile soils, sparse vegetation, and intense episodic rainfall can generate severe but spatially uneven soil loss. These processes reduce soil productivity, increase sediment movement, degrade downstream water quality, and accelerate desertification (Walling, 2017; Panagos et al., 2022; Rashmi et al., 2022; Firoozi and Firoozi, 2024). In arid and hyper-arid watersheds, erosion is often concentrated in limited zones controlled by topography, drainage convergence, vegetation cover, and soil erodibility. Therefore, spatially explicit erosion modeling is essential for identifying priority areas for conservation and watershed management.

In Saudi Arabia, erosion risk is intensified by the Kingdom’s predominantly hyper-arid climate, low rainfall, sparse vegetation, and weakly developed soils. Although annual rainfall is generally low, short-duration storms can generate high runoff and localized erosion, especially in wadis and drainage corridors where flow is concentrated (Odnoletkova and Patzek, 2023; Hassani et al., 2020; Ghazanfar, 2024; Karimi et al., 2025). National initiatives such as the Saudi Green Initiative and recent land-restoration efforts highlight the importance of controlling land degradation. However, the spatial distribution of soil loss and sediment delivery remains poorly quantified in many interior watersheds of Saudi Arabia, particularly in hyper-arid basins where field measurements are limited (Bahrawi et al., 2016; Kousar and Shirazi, 2023).

The Revised Universal Soil Loss Equation (RUSLE) remains one of the most widely used empirical models for estimating potential soil loss because it can integrate rainfall erosivity, soil erodibility, slope length and steepness, vegetation cover, and conservation-practice factors within a GIS framework (Renard et al., 1997; Panagos et al., 2015). Recent studies have improved RUSLE-based erosion assessment by combining it with remote sensing, high-resolution digital elevation models, Sediment Delivery Ratio (SDR) approaches, and machine learning. RUSLE–SDR methods are particularly useful because they distinguish gross soil loss from the fraction of sediment likely to reach the watershed outlet (Halder, 2023; Ghimire et al., 2024; Yang et al., 2024). At the same time, machine learning models can support internal consistency assessment and sensitivity analysis by identifying non-linear relationships among erosion-controlling factors (Filchev and Kolev, 2023; Alruzuq and Mossa, 2025; Sahour et al., 2021; Rahaman et al., 2025; Ge et al., 2023; Southworth et al., 2024).

Despite these advances, most integrated RUSLE applications remain concentrated in humid, semi-humid, or non-Saudi environments. In dryland and data-limited settings, recent studies show the value of combining GIS-based erosion modeling with machine learning and remote sensing to improve erosion prediction and susceptibility mapping (Gueffaf et al., 2026; Baiddah et al., 2023; Sahu et al., 2025; Alruzuq, 2026). In Saudi Arabia, RUSLE-based studies have been conducted using satellite rainfall, soil, vegetation, and DEM data; however, many applications stop at gross erosion estimation and do not explicitly assess sediment delivery or compare modeled erosion patterns with supporting surface-condition indicators.

For example, Ejaz et al. (2023) estimated soil loss in Wadi Baysh using CHIRPS rainfall, soil data, Sentinel-2 imagery, and ALOS PALSAR DEM, while Alsaihani and Alharbi (2024) applied RUSLE to Wadi Bin Abdullah. Other work has used machine learning to assess soil erodibility, but without linking predicted erosion to sediment yield or surface thermal–vegetation conditions (Alqadhi et al., 2023). This creates a methodological gap in Saudi Arabia: limited integration of RUSLE with SDR, machine learning-based sensitivity assessment, and remote-sensing surface indicators such as SERI.

This study addresses this gap by developing an integrated soil erosion assessment for a delineated hyper-arid study watershed containing Wadi Samnan near Az Zulfi, central Saudi Arabia. The central model is RUSLE, which estimates spatial soil loss from rainfall erosivity, soil erodibility, topography, vegetation cover, and support-practice factors. SDR is then used to estimate sediment delivery, SERI is used as a remote-sensing surface-condition check based on NDVI and land surface temperature, and Random Forest is applied as an internal consistency and sensitivity-analysis tool. Together, these components provide a focused framework for identifying erosion hotspots, sediment-contributing zones, and priority conservation areas in the study watershed. The specific objectives are to: (1) map spatial patterns of soil erosion using RUSLE; (2) estimate sediment delivery and local sediment retention using SDR; (3) evaluate erosion-prone surfaces using SERI; and (4) assess the relative importance of erosion-controlling factors using Random Forest.

2 Study area

The study watershed, which contains Wadi Samnan near Az Zulfi, is located in Riyadh Province, central Saudi Arabia, within the Najd Plateau and near the prominent Tuwaiq (Tweig) escarpment (Figure 1). In this study, “the study watershed” refers to the delineated drainage polygon used for the erosion and sediment-delivery analysis, whereas “Wadi Samnan” refers to the named wadi located within this watershed. The surrounding physiographic setting is influenced by the Tuwaiq escarpment, a prominent plateau-margin feature that rises about 600 m above the surrounding terrain and exposes Middle to Upper Jurassic stratigraphy, including Arab-D equivalents used as analogs for subsurface reservoirs (Sen et al., 2023). The region has a hot desert climate (Köppen BWh), characterized by extremely hot, arid summers, mild winters, and low but spatially variable rainfall. To avoid mixing general regional climate descriptions with the rainfall input used in the RUSLE calculation, precipitation in this study is reported using the CHIRPS-derived mean annual precipitation for 2019–2024, which ranged from 138.67 to 162.69 mm yr^-1 across the study watershed. Rainfall events are episodic and may occur as short-duration storms capable of producing localized runoff and erosion along wadi channels and drainage corridors. Geologically, the area is underlain by a series of Mesozoic carbonate formations, especially the Late Jurassic to Cretaceous Sulaiy and Yamama limestones. In some places, these are overlain by the Arab and Hith Formations, which include anhydrite. Quaternary alluvial and aeolian deposits fill the wadis and valleys (Al-Hashim et al., 2024). These rocks create karst features and collapse structures, especially along escarpments, and host both shallow alluvial aquifers and deeper carbonate aquifers, such as the Sulaiy/Yamama, which produce groundwater through fractures and weathered zones. Overall, the study area features arid landforms, an aridic soil-moisture regime, and complex geology, which together make its hydrological and environmental dynamics important for understanding regional water resources and landscape change.

FIGURE 1

3 Methodology

3.1 Data used

This study used geospatial datasets from 2019 to 2024 to analyze soil erosion in the study watershed. All datasets were filtered by date to cover the specified period, confined by the watershed boundary, and clipped to the study area. Sentinel-2 Surface Reflectance imagery from Copernicus was retrieved via Google Earth Engine (GEE) (Table 1). Landsat 8 Collection 2 Tier 1 Top-of-Atmosphere (TOA) reflectance and brightness-temperature data from the OLI/TIRS sensors, provided by the United States Geological Survey (USGS), were accessed through Google Earth Engine. CHIRPS Pentad precipitation data, produced by the Climate Hazards Group at the University of California, Santa Barbara (UCSB/CHG), was used for rainfall input. Soil texture data from EnvirometriX Ltd/OpenLandMap was obtained from the OpenLandMap dataset via GEE (Figure 2). Land-use/land-cover information used for P-factor assignment was obtained from Google Dynamic World V1 through Google Earth Engine. Dynamic World is a near-real-time global 10 m LULC product generated from Sentinel-2 imagery using a pretrained Fully Convolutional Neural Network and provides per-pixel class probabilities and discrete land-cover labels. Topographic information was derived from the ALOS PALSAR Digital Elevation Model (DEM) with a 12.5-m resolution, downloaded from the Alaska Satellite Facility (ASF).

FIGURE 2

All GIS and remote sensing analyses were conducted using GEE for satellite data preprocessing, vegetation index generation, rainfall processing, and land-cover data extraction. ArcGIS Pro 3.6.3 was used for GIS-based spatial analysis, watershed processing, raster calculations, map production, and layout design. Python 3.x was used for statistical analysis, correlation assessment, and Random Forest sensitivity analysis. Specifically, NDVI, LST, rainfall-derived layers, and land-cover-related rasters were generated in GEE, while DEM-derived layers, including slope, flow direction, flow accumulation, watershed boundary, drainage network, and LS factor, were produced in ArcGIS Pro 3.6.3 using Spatial Analyst and Hydrology tools. SDR and sediment-yield rasters were then calculated in ArcGIS Pro 3.6.3 using Raster Calculator based on the derived slope, watershed area, and RUSLE soil-loss layers. Python 3.x was used for Pearson correlation analysis, internal consistency assessment, and Random Forest sensitivity analysis. Lastly, all numerical values, class ranges, percentages, and statistical outputs reported in the Results section, figures, and tables were checked against the final raster outputs and statistical analysis results to ensure consistency across the manuscript.

3.2 RUSLE model

The RUSLE was used to estimate annual water-induced soil loss across the study area. Although RUSLE was originally developed for agricultural hillslopes, it is widely applied in GIS-based erosion studies in arid and semi-arid watersheds because it provides a practical framework for integrating rainfall erosivity, soil erodibility, topography, vegetation cover, and support-practice conditions where long-term field erosion measurements are unavailable. In this study, RUSLE was used as an empirical spatial screening tool to identify relative patterns of rainfall–runoff-driven soil-loss potential rather than as a direct measurement of total erosion. The model is expressed mathematically as:Where is the estimated annual soil loss (t ha^-1 yr^-1), is the rainfall-runoff erosivity factor, is the soil erodibility factor, is the slope length and steepness factor, is the cover-management factor, and is the support-practice factor. The RUSLE factors were derived from the datasets summarized in Table 1 and processed in a GIS environment, with R derived from CHIRPS rainfall data, K from OpenLandMap soil texture, LS from the ALOS PALSAR DEM, C from Sentinel-2-derived NDVI, and P from Google Dynamic World V1 land-use/land-cover data combined with slope-based proxy values. Furthermore, it should be noted that RUSLE estimates water-induced soil erosion driven mainly by rainfall, runoff, soil erodibility, topography, vegetation cover, and support-practice conditions. Therefore, the resulting erosion maps represent RUSLE-based water erosion risk rather than total erosion risk, and they do not explicitly quantify wind-driven aeolian erosion.

3.2.1 R factor (rainfall-runoff erosivity)

The rainfall-runoff erosivity factor (R) was calculated using CHIRPS precipitation data for the 2019–2024 study period. The CHIRPS precipitation data were filtered to the study period, clipped to the watershed boundary, and processed to generate a mean annual precipitation raster. The annual mean precipitation raster was then converted into point features, interpolated using Inverse Distance Weighting (IDW), and used to derive the rainfall-runoff erosivity layer. To maintain consistency between the climate description, Figure 3a, and the R-factor calculation, the precipitation term used in the R-factor equation refers specifically to the CHIRPS-derived mean annual precipitation for 2019–2024, not to a separate long-term regional climatological average.

FIGURE 3

The erosivity index was calculated using the precipitation-based equation applied by Ejaz et al. (2023) for Wadi Baysh, Saudi Arabia. This equation was selected because rainfall-intensity data were unavailable for the study watershed and because precipitation-based R-factor equations are commonly used in data-limited arid-region RUSLE applications. However, the equation was not assumed to be the most accurate formulation for all arid environments. Instead, it was used as a regional empirical approximation that is transferable with caution because both Wadi Baysh and the present study area are Saudi Arabian dryland wadi systems where rainfall is episodic and erosion is commonly driven by short-duration runoff events. At the same time, differences in rainfall regime, storm intensity, and local physiography mean that the R-factor results should be interpreted as RUSLE-based screening estimates rather than field-calibrated erosivity measurements. Future work should compare alternative R-factor equations and calibrate rainfall erosivity using local rain-gauge or rainfall-intensity data when such observations become available.

The R-factor was calculated as:where is the rainfall-runoff erosivity factor and is the CHIRPS-derived mean annual precipitation for 2019–2024 in mm yr^-1. Therefore, the R-factor raster directly reflects the spatial precipitation pattern shown in Figure 3a.

3.2.2 K-factor (soil erodibility)

The soil erodibility factor (K) was determined using soil texture classification data from OpenLandMap. Because measured soil properties such as organic matter, soil structure, permeability, and detailed particle-size distribution were not available for the study watershed, representative RUSLE K values were assigned based on USDA soil texture classes. The K-factor values are reported in SI RUSLE units of t ha h ha^-1 MJ^-1 mm^-1. Areas containing waterbodies or lacking valid soil texture data were assigned K = 0.00. Sand was assigned K = 0.25, Loamy Sand was assigned K = 0.35, Sandy Clay Loam was assigned K = 0.35, and Sandy Loam was assigned K = 0.45.

Accordingly, Sandy Loam represents the most erodible texture class in the model because its greater fine-particle content and lower aggregate stability generally increase susceptibility to detachment compared with coarser sandy materials. Sand is treated as a moderately erodible class, while Loamy Sand and Sandy Clay Loam are treated as moderately high erodibility classes. These representative values follow standard texture-based RUSLE parameterization practices used when detailed soil physical measurements are unavailable (Renard et al., 1997; Corral-Pazos-de-Provens et al., 2023). To avoid scaling confusion, all K-factor values are reported using the same unit system throughout the Methods, Results, Figure 4b, and Discussion.

FIGURE 4

3.2.3 LS factor (slope length and steepness)

The topographic factor (LS) was derived from slope degree data calculated from the DEM, combining both slope length and slope steepness as described in the RUSLE framework (Renard et al., 1997; Moore and Burch, 1986; Schmidt et al., 2019). The slope steepness was calculated in radians from the DEM-derived slope map, and the slope length (λ) was estimated based on the relationship between cell size and the sine of the slope angle.

This method considers both the slope gradient and locally derived slope length in erosion processes, while avoiding potential distortions caused by flow accumulation. A maximum slope length of 200 m was set to prevent unrealistically high LS values in areas with very low slopes.

3.2.4 C-factor (cover management)

The C-factor was derived from the NDVI calculated using Sentinel-2 imagery. A composite mean of Sentinel-2 data from 2019 to 2024 was computed, using the relevant spectral bands to ensure temporal consistency. NDVI was calculated using the standard formula:Where B8 and B4 represent Sentinel-2’s near-infrared and red bands, respectively.

This formulation captures how vegetation cover reduces soil exposure to erosive forces, with higher NDVI values yielding lower C-factor values and indicating stronger vegetation protection. To maintain consistency with the standard RUSLE interpretation, the derived C-factor raster was constrained to the 0–1 range before the RUSLE calculation. The corrected C-factor values ranged from 0.38 to 1.00, with values near 1 representing bare or poorly protected surfaces and lower values indicating stronger vegetation protection. The corrected C-factor raster was subsequently used to regenerate the final RUSLE soil-loss raster and all dependent sediment-yield, class-area, correlation, Random Forest, figure, and table outputs.

3.2.5 P-factor (support practice factor)

The support practice factor (P-factor) in the RUSLE measures how well conservation practices, such as contouring, strip-cropping, and terracing, reduce soil erosion compared to a standard plot without these practices (P = 1) (Panagos et al., 2015). Because spatial data on actual conservation measures are often limited, most studies use a proxy approach, estimating P-values based on land-use/land-cover (LULC) and slope categories. This method enables more spatially explicit modeling by assigning P-values that reflect the observed effectiveness of erosion control.

3.2.5.1 Land use/land cover (LULC) classification

Land-use/land-cover (LULC) information used to support P-factor assignment was obtained from Google Dynamic World V1 (Brown et al., 2022; Table 1), a near-real-time global 10 m land-cover product accessed through Google Earth Engine. Dynamic World applies a pretrained Fully Convolutional Neural Network (FCNN) to individual Sentinel-2 images and provides per-pixel probabilities and corresponding discrete land-cover labels. The FCNN was developed and trained by the Dynamic World data producers using globally distributed annotated Sentinel-2 imagery. Its single-image land-cover outputs achieved an overall agreement of 73.8% with expert-consensus reference annotations (Brown et al., 2022). Because Dynamic World is an externally developed, pretrained, and published product, no local CNN training, validation, or testing was conducted in this study. The published agreement provides an external and verifiable product-level quality benchmark for the LULC input; however, it should not be interpreted as a site-specific accuracy assessment for the Wadi Samnan watershed. Additionally, the Dynamic World image collection was filtered to the 2019–2024 study period and the watershed boundary, and the pixel-wise temporal mode of the discrete label band was calculated to generate the representative LULC layer used for P-factor assignment.

The Dynamic World classes present in the study area—bare ground, trees, crops, shrub and scrub, grass, built area, water, and flooded vegetation—were reclassified into the land-cover categories used in this study: Bare/Sparse Vegetation, Tree Cover, Cropland, Shrubland, Grassland, Built-Up, Permanent Water Bodies, and Herbaceous Wetland. These classes were mapped for the study watershed in Figure 5a and used to assign the corresponding P-factor proxy values described in Section 3.2.5.2. The Dynamic World-derived LULC layer was used only to support P-factor assignment within the RUSLE framework and was not treated as an independently developed erosion model or as a separate analytical objective.

FIGURE 5

3.2.5.2 P-factor value computation

P-factor values were assigned as follows: water bodies received P = 0.0 to indicate complete protection; herbaceous wetland received P = 0.1, reflecting near-complete surface protection under saturated, low-erosivity conditions; built-up areas and cropland, shrubland, and grassland received P = 0.5, reflecting limited erosion protection in the absence of formal conservation measures; tree-covered areas were assigned a fixed P = 0.4 to represent the higher erosion protection provided by canopy interception, ground cover, and root stabilization; and bare/sparse vegetation, the dominant land-cover class, was given slope-dependent values ranging from P = 0.75 on slopes less than 2% to P = 0.9 on slopes 20% or greater, recognizing that as slope increases, runoff speed increases and conservation effectiveness decreases. These slope-dependent values follow the guidelines outlined in Agricultural Handbook No. 703 and align with the ranges used in recent large-scale P-factor modeling studies (Panagos et al., 2015; Ebabu et al., 2022; Yaswanth et al., 2022; Thapa, 2020). The complete LULC- and slope-based P-factor assignments applied in this study are summarized in Table 2.

3.3 Sediment delivery ratio (SDR)

To estimate sediment-delivery potential, the Sediment Delivery Ratio (SDR) was calculated using an empirical formulation adapted from Vanoni (1975). Vanoni’s SDR equation is traditionally a lumped basin-scale relationship; therefore, in this study it was not used as a calibrated measurement of observed sediment delivery. Instead, it was spatialized in GIS as a relative sediment-delivery index to show how sediment-transfer potential varies across the study watershed. This type of empirical SDR approach is commonly used in watershed-scale sediment studies where direct sediment transport measurements are unavailable (Kim and Julien, 2006).

The SDR was calculated as:Where A is the total delineated watershed area in km² and was kept constant across the raster calculation, while Slope is the local DEM-derived slope (%) at each pixel. Thus, the spatial variation in the SDR map is controlled by local slope differences, whereas the watershed-area term represents the overall basin-size effect. The SDR raster was generated in ArcGIS Pro 3.6.3 using Raster Calculator by applying the equation to the local slope raster while using the total watershed area as a constant parameter.

This approach provides a spatially distributed approximation of sediment-delivery potential, allowing areas with steeper slopes and stronger topographic connectivity to be distinguished from low-slope areas where sediment retention is more likely. Because the original Vanoni relationship is lumped and empirical, the resulting SDR map should be interpreted as a relative sediment-delivery potential map rather than a field-calibrated estimate of actual sediment export. This limitation is particularly important in hyper-arid wadis, where episodic runoff, channel storage, and sediment retention can strongly influence actual sediment movement.

3.4 Sediment yield

Sediment yield rate was calculated as the SDR-adjusted fraction of RUSLE-estimated soil loss. At the pixel scale, the area-normalized sediment-yield rate was estimated as:Where is the RUSLE-estimated water-induced soil loss at pixel i in t ha^-1 yr^-1, and is the spatialized sediment-delivery ratio at the same pixel. Because is dimensionless, is also expressed in t ha^-1 yr^-1. The resulting sediment-yield raster therefore represents relative spatial patterns of potential sediment delivery from each pixel, not measured total sediment mass at the watershed outlet.

3.5 Surface erosion risk index (SERI)

SERI was derived by combining satellite-based NDVI and Land Surface Temperature (LST) layers to highlight areas of exposed soil prone to erosion. A time series of NDVI and LST was computed using Landsat 8 Collection 2 Tier 1 TOA reflectance and Band 10 brightness-temperature data for the period 2019–2024. The mean of each index’s time series was computed, and the resulting mean NDVI and mean LST over this period were then used to calculate the SERI.

NDVI was calculated based on the near-infrared (NIR, Band 5) and red (Band 4) reflectance using the standard formula:

NDVI values were then normalized across the study area to derive the fractional vegetation cover (FV) using:

Surface emissivity (ε) was estimated from FV using the empirical relationship:

To compute LST in degrees Celsius, the thermal band (Band 10) was used as brightness temperature and integrated with emissivity using the radiative transfer equation:where λ = 11.5 μm (1.15 × 10⁻⁵ m) is the effective wavelength used for Landsat 8 TIRS Band 10, and ρ = 1.438 × 10⁻² m K is the second radiation constant.

Using these, SERI was calculated as:

SERI is defined in this study as a relative index of vegetation deficiency scaled by Land Surface Temperature. Because LST occurs in the denominator, an increase in LST alone does not necessarily produce a higher SERI value. Instead, higher SERI values primarily indicate greater vegetation deficiency relative to the corresponding surface temperature. The index is therefore used as a supporting relative surface-condition and erosion-susceptibility proxy rather than as an independently calibrated or field-validated erosion measure. Low vegetation cover and elevated surface temperatures are commonly associated with exposed and environmentally stressed dryland surfaces (Anand et al., 2025; Adou et al., 2024).

3.6 Random forest internal consistency and sensitivity analysis

Random Forest (RF) was used as an internal consistency and sensitivity analysis tool rather than as an independent ground-validation method. The RUSLE soil-loss map was used as the response variable, while the RUSLE factors (R, K, LS, C, and P) were used as predictors. Valid raster pixels were randomly divided into two subsets: 80% of the pixels were used as the training subset to fit the RF model, and the remaining 20% were retained as a held-out test subset for evaluating model performance. No separate validation subset was used for hyperparameter tuning because the RF analysis was not intended to develop an operational predictive model; instead, it was used to examine whether the spatial pattern of RUSLE-estimated soil loss could be reproduced from the input factors and to identify the relative influence of each erosion-controlling variable. Lastly, model performance on the held-out test subset was evaluated using R², RMSE, and MAE, while feature importance was used to interpret the contribution of each factor. Because the RF model was trained using RUSLE-derived outputs, the results are interpreted as an internal model-consistency and sensitivity check, not as independent validation against observed erosion.

4 Results

4.1 RUSLE factor results

4.1.1 Rainfall erosivity factor (R)

The CHIRPS-derived mean annual precipitation used in the RUSLE calculation shows spatial variation across the study watershed, ranging from ≤138.67 mm yr^-1 in the southern regions to >162.69 mm yr^-1 in the northern areas for the period 2019–2024 (Figure 3a). These precipitation values represent the study-period rainfall input used to derive the R-factor and therefore provide the basis for the rainfall-runoff erosivity pattern shown in Figure 3b. Using the precipitation-based erosivity equation, the mean R-factor values ranged from ≤69.81 to >83.31 MJ mm ha^-1 h^-1 yr^-1 throughout the watershed. The distribution indicates that the northern parts of the watershed experience higher rainfall erosivity than the southern parts. The highest erosivity category (>83.31 MJ mm ha^-1 h^-1 yr^-1) occurs mainly in the northern and northwestern sections, whereas the lowest values (≤69.81 MJ mm ha^-1 h^-1 yr^-1) occur in the southern and southeastern areas. Overall, the highest R-factor class is approximately 19% higher than the lowest, indicating a clear but moderate north–south gradient in rainfall-driven erosion potential.

4.1.2 Soil erodibility factor (K)

The study area mainly consists of four mapped soil-texture classes covering a total of 291,175 ha. Sand is the dominant class, covering 236,556.25 ha, or approximately 81.2% of the mapped soil-texture area. Loamy Sand covers 27,975 ha, Sandy Loam covers 25,281.25 ha, and Sandy Clay Loam covers 1,362.5 ha (Figure 4a). The assigned K-factor values range from 0.00 for waterbodies or areas without valid soil texture data to 0.45 t ha h ha^-1 MJ^-1 mm^-1 for the most erodible soil class (Figure 4b). Among the mapped soil texture classes, Sandy Loam has the highest assigned erodibility value (K = 0.45), followed by Loamy Sand and Sandy Clay Loam (K = 0.35), while Sand has a moderate assigned erodibility value (K = 0.25).

Therefore, although Sand is the most spatially dominant soil texture in the watershed, it is not the highest K-factor class in the adopted RUSLE parameterization. The highest soil erodibility is associated with Sandy Loam areas because their relatively higher fine-particle content increases susceptibility to detachment compared with coarser sandy materials. The final K-factor map, therefore, reflects both the spatial dominance of sandy surfaces and the higher erodibility assigned to Sandy Loam patches. All K-factor values are reported consistently in SI RUSLE units of t ha h ha^-1 MJ^-1 mm^-1.

4.1.3 Slope length and steepness factor (LS)

The topographic analysis shows notable spatial differences in erosion potential throughout the watershed, with elevation ranging from 629 m to 829 m and displaying moderate relief that produces varied erosion conditions (Figure 6). The slope analysis indicates that much of the watershed is covered by gentle to moderate slopes, especially the 0°–1.18° and 1.19°–4.13° classes, which generally correspond to low to moderate slope-induced erosion potential. In contrast, steeper slopes, particularly the 14.94°–50.1° class, are mainly located in the western and northwestern parts of the watershed and form localized zones of higher erosion susceptibility.

FIGURE 6

The LS-factor map was classified using the same class ranges shown in Figure 6b: 0, 0.01–2.03, 2.04–3.82, 3.83–5.79, and 5.80–15.23. Most of the watershed is characterized by low LS values between 0 and 3.82, indicating that the prevailing topographic conditions generally limit slope-length and slope-steepness effects on erosion. Higher LS values of 3.83–5.79 and 5.80–15.23 occur mainly along steeper western slopes and drainage-related terrain, where slope steepness and flow convergence increase runoff erosivity and soil-loss potential. Therefore, the spatial pattern of the LS factor indicates that topographic control on erosion is generally limited across broad low-gradient areas but becomes important in localized steep-slope and drainage-corridor zones.

4.1.4 Cover-management factor (C)

NDVI values range from −0.09 to 0.58, indicating generally sparse vegetation across the watershed (Figure 7a). The derived C-factor values range from 0.38 to 1.00, with high values dominating the central, eastern, and southern areas where vegetation cover is limited (Figure 7b). Lower C-factor values occur mainly in scattered patches along drainage corridors and topographically favorable microsites, where vegetation provides localized protection against soil erosion.

FIGURE 7

4.1.5 Support-practice factor (P)

The LULC layer used for P-factor assignment was derived from Google Dynamic World V1 and reclassified into eight study categories: Bare/Sparse Vegetation, Tree Cover, Cropland, Shrubland, Grassland, Built-Up, Permanent Water Bodies, and Herbaceous Wetland. The P-factor (support practice factor) ranges from 0.0 to 0.9 across the watershed, with its spatial distribution reflecting the combined effects of land-use types and slope conditions (Figure 5b). The dominant land-cover class is Bare/Sparse Vegetation, covering 280,688.63 ha, or approximately 96.4% of the valid classified LULC area. It has slope-dependent P-values ranging from 0.75 to 0.90, depending on terrain steepness. Tree-covered areas (445.37 ha) are assigned a P value of 0.4, indicating better erosion protection because of the forest canopy and roots. Cropland (4,581.69 ha), Shrubland (163.51 ha), and Grassland (311.49 ha) are given moderate P values of 0.5. Built-up areas covering 4,918.04 ha also have P = 0.5, recognizing limited erosion risk on impervious surfaces. Small areas of Permanent Water Bodies (15.37 ha) and Herbaceous Wetland (0.16 ha) are assigned very low P values of 0.0 and 0.1, respectively. Overall, the high P-factor values across most of the watershed suggest limited use of conservation practices, with erosion control mainly relying on natural land cover rather than engineered measures.

4.1.6 RUSLE-based soil loss results

The RUSLE model results (Figures 8, 9) reveal significant spatial variation in water-induced soil-loss potential across the watershed, with values ranging from 0 to 319.5 t ha^-1 yr^-1. To maintain consistency between the map, chart, and text, soil-loss values were classified into five severity classes: slight erosion (<10 t ha^-1 yr^-1), moderate erosion (10–20 t ha^-1 yr^-1), high erosion (20–30 t ha^-1 yr^-1), very high erosion (30–40 t ha^-1 yr^-1), and severe erosion (>40 t ha^-1 yr^-1). The slight erosion category dominates the watershed, covering 43.2% of the total area, indicating generally stable soil conditions across much of the landscape. Moderate erosion affects 28.3% of the watershed and represents transitional zones between stable areas and higher-risk terrain, likely associated with moderate slopes, agricultural fields, or areas with partial vegetation cover. High erosion affects 11.2% of the area, while very high erosion occurs across 6.4% of the watershed, representing zones where steeper slopes, land-use conditions, or reduced vegetation cover contribute to elevated soil loss. Severe erosion areas, exceeding 40 t ha^-1 yr^-1 and reaching a maximum of 319.5 t ha^-1 yr^-1, cover 10.9% of the watershed and occur mainly along drainage networks in the western and central parts of the study area. These severe erosion hotspots represent the primary zones of soil degradation and should be prioritized for conservation measures to reduce continued land degradation and sediment transport.

FIGURE 8

FIGURE 9

4.2 Sediment delivery ratio and sediment yield

The SDR values across the watershed range from 0.009 to 0.151, with most areas showing very low delivery ratios of 0.009–0.013 (Figure 10a). This spatial pattern indicates that most of the watershed has a limited ability to transport eroded sediment to the outlet, suggesting that much of the detached soil is likely deposited locally rather than transported downstream. Because SDR represents the fraction of eroded material potentially delivered, these values imply local sediment-retention potential of approximately 85%–99% across the watershed. Specifically, the highest SDR value of 0.151 corresponds to approximately 84.9% retention, while the lowest SDR value of 0.009 corresponds to approximately 99.1% retention. The higher SDR values of 0.041–0.151 are concentrated along the main drainage channels and steep-slope areas, especially along the central north–south drainage network, where concentrated flow paths and steeper terrain increase sediment-transfer potential. The overall low SDR values across most of the watershed, calculated using Vanoni’s equation with basin area and local slope effects, suggest that watershed topography and basin configuration promote substantial internal sediment retention, with only a limited fraction of gross erosion likely to reach the basin outlet.

FIGURE 10

The sediment-yield rate (SY) results show a notably different spatial pattern from erosion potential, with values ranging from 0 to 39.18 t ha^-1 yr^-1 (Figure 10b), underscoring the crucial role of sediment delivery processes in controlling watershed sediment export. About 95% of the watershed exhibits very low modeled sediment-yield rates (0–0.46 t ha^-1 yr^-1), indicating that, despite localized high erosion rates, the potential transfer of eroded material toward the watershed outlet is strongly limited by topographic and hydrological factors that promote sediment deposition within the basin. The higher sediment-yield-rate categories (5.08–39.18 t ha^-1 yr^-1) are confined to less than 1% of the watershed area; however, these small zones represent the main areas where high erosion rates and efficient transport pathways coincide. This contrast between widespread erosion hotspots and limited high-yield zones suggests low overall sediment connectivity, consistent with the SDR-derived retention estimate that approximately 85%–99% of eroded material may be retained within the watershed through hillslope deposition, local storage, or channel retention. The maximum sediment-yield rates, approaching 39.18 t ha^-1 yr^-1, occur in critical areas where the erosion–transport–delivery process operates at peak efficiency, likely along steep channel banks, active gully systems, or direct sediment pathways to streams. This highlights the importance of targeting conservation efforts on these disproportionately influential zones.

4.3 Surface erosion risk index (SERI)

The SERI values range from 0.0116 to 0.0414 across the study watershed (Figure 11). The index represents vegetation deficiency relative to Land Surface Temperature, with higher values indicating greater vegetation deficiency after thermal scaling. Approximately 70% of the watershed has SERI values above 0.0249 and therefore falls within the higher relative SERI classes. These areas are mainly concentrated in the eastern and central parts of the watershed, whereas lower values occur in smaller areas with comparatively lower vegetation deficiency relative to surface temperature. The SERI classes represent relative erosion-susceptibility rankings within the study watershed and should not be interpreted as absolute erosion-risk probabilities or as evidence that increasing LST alone necessarily increases SERI.

FIGURE 11

4.4 Internal consistency assessment

The Pearson correlation results are summarized in Table 3. The relationship between RUSLE and SERI is weak but positive (r = 0.120), indicating a weak association between RUSLE-estimated soil loss and the SERI vegetation-deficiency-to-temperature ratio (Figure 12a). Because LST occurs in the denominator of the SERI formulation, this relationship should not be interpreted as showing that higher temperature alone increases SERI. Instead, SERI is treated as a supporting relative surface-condition indicator rather than as independent validation of measured erosion.

FIGURE 12

The relationship between RUSLE and SDR is strongly positive (r = 0.898), indicating close internal consistency between modeled soil-loss patterns and sediment-delivery potential (Figure 12b). However, this strong relationship should be interpreted cautiously because RUSLE and SDR share topographic controls, particularly slope. The reported p-values are expressed as p < 0.001 rather than p = 0. Because the analysis uses a very large number of spatially autocorrelated pixels, statistical significance is expected and is not the primary basis for interpretation. Instead, the Pearson r values are used to interpret the strength and direction of the relationships. Thus, the RUSLE–SERI and RUSLE–SDR comparisons are treated as internal consistency checks, not as independent ground validation.

4.5 Random forest sensitivity analysis

To assess the internal consistency and sensitivity of the RUSLE framework, a Random Forest regression was applied using the RUSLE factors (K, R, C, LS, and P) as predictors and the computed RUSLE soil-loss values as the response variable (Figure 13). The RF model was fitted using the 80% training subset and evaluated using the 20% held-out test subset. Performance evaluation on the held-out test subset yielded an RMSE of 12.8305, MAE of 6.9436, and R² of 0.8173, indicating that the input factors can reproduce much of the spatial variability in the RUSLE-derived soil-loss estimates. Because the response variable was mathematically derived from the same five RUSLE factors used as predictors, the R² of 0.8173 largely reflects the Random Forest model re-learning the relationships embedded in the RUSLE equation. Accordingly, these statistics are interpreted as an internal consistency check rather than independent validation or prediction of observed erosion on the ground. Feature importance analysis was therefore used mainly to identify the relative sensitivity of the RUSLE output to its input factors. The results show that LS and the land-cover- and slope-based P-factor proxy contributed most strongly, followed by R, C, and K, indicating that topographic conditions and support-practice assumptions had the greatest influence on modeled soil-loss variability.

FIGURE 13

4.6 NDVI time-series analysis

The NDVI time series analysis from 2019 to 2024 reveals distinct temporal and spatial patterns in vegetation cover, with significant implications for erosion susceptibility across the watershed (Figure 14). The data show consistent spatial distribution patterns throughout the study period, with higher vegetation density concentrated along the central drainage network and scattered patches in the northern and western portions of the watershed. At the same time, the majority of the area maintains consistently low vegetation cover across all years. All six annual NDVI maps were displayed using a common classification: ≤0.00, 0.01–0.10, 0.11–0.20, 0.21–0.30, and >0.30. This consistent classification shows that low NDVI values dominated most of the watershed throughout 2019–2024, whereas localized higher values occurred mainly along drainage corridors and scattered vegetated patches. The temporal variations suggest that, although the overall spatial pattern remains relatively stable, there are subtle year-to-year fluctuations in vegetation vigor throughout the study period. The persistent dominance of low NDVI values across most of the watershed throughout the time series confirms the arid nature of the environment. These annual patterns suggest that vegetation-related erosion protection remained broadly limited from 2019 to 2024, with localized improvements along drainage corridors providing limited but important surface protection. A common NDVI classification was applied to all six annual panels to support direct visual comparison of spatial vegetation patterns across the 2019–2024 period.

FIGURE 14

4.7 Land surface temperature time-series analysis

The Land Surface Temperature (LST) time series from 2019 to 2024 reveals distinct temporal and spatial thermal patterns across the watershed, providing important insights into surface conditions that affect erosion susceptibility (Figure 15) (Sahu et al., 2024). The analysis reveals consistent spatial distribution, with cooler temperatures concentrated along the central drainage corridor and scattered areas in the northern regions, while the majority of the watershed experiences elevated surface temperatures across all study years. All six annual LST maps were displayed using a common classification: ≤30.00 °C, 30.01 °C–32.50 °C, 32.51 °C–35.00 °C, 35.01 °C–37.50 °C, and >37.50 °C. Using these fixed classes, 2021 shows the largest extent of the highest-temperature class, whereas 2022 shows a greater extent of the cooler classes. The persistent high-temperature zones, particularly in the western, southern, and eastern portions of the watershed, indicate exposed soil and rock surfaces with minimal thermal buffering from vegetation or moisture. The cooler thermal signatures along drainage networks likely reflect the influence of topographic shading, occasional moisture retention, and slightly higher vegetation density in these areas. These thermal patterns reveal significant spatial heterogeneity in surface energy balance conditions, with extensive high-temperature zones indicating widespread exposure of thermally active bare surfaces that are potentially more susceptible to physical weathering processes, which can contribute to soil detachment and surface degradation. A common LST classification was applied to all six annual panels to support direct visual comparison of spatial thermal patterns across the 2019–2024 period.

FIGURE 15

5 Discussion

The integrated erosion assessment framework applied in this study provides a spatially explicit interpretation of water-induced soil-loss potential, sediment-delivery behavior, and surface vulnerability in a hyper-arid watershed containing Wadi Samnan near Az Zulfi, Saudi Arabia. By combining RUSLE, SDR, remote sensing indicators, and Random Forest-based internal sensitivity analysis, the study distinguishes between areas where soil detachment is likely to occur and areas where eroded sediment is more likely to be transferred through the drainage system. This distinction is particularly important in dryland wadis, where erosion processes are spatially concentrated and episodic, while sediment delivery is often constrained by local deposition, channel storage, and weak hydrological connectivity. As a result, the framework provides a practical basis for identifying erosion hotspots, sediment-contributing zones, and priority management areas in a data-scarce environment.

The rainfall erosivity results indicate a clear but moderate spatial gradient across the watershed, with the highest R-factor class approximately 19% higher than the lowest class. Although mean annual rainfall remains low in central Saudi Arabia, short-duration, high-intensity rainfall events can generate concentrated runoff that triggers localized erosion, particularly along drainage corridors and steep slopes. In hyper-arid systems, geomorphic change is therefore often controlled less by average rainfall and more by episodic storm events that briefly trigger runoff and sediment transport. This issue is especially relevant under projected climate variability, as recent studies suggest that extreme precipitation and flood-generating events may intensify across parts of the Arabian Peninsula, even where long-term mean rainfall remains low or variable (Panagos et al., 2022; Patlakas et al., 2021; Karimi et al., 2025). Under such conditions, areas where rainfall erosivity coincides with steep slopes, sparse vegetation, exposed surfaces, and erodible soils may become increasingly vulnerable to accelerated erosion. Therefore, the mapped hotspots in this study are relevant not only to present-day erosion assessment but also to climate-adaptation planning.

Soil erodibility further reinforces this spatial vulnerability. The mapped soil-texture area is dominated by sandy surface materials, with Sand covering approximately 81.2%. However, in the adopted RUSLE parameterization, Sandy Loam represents the most erodible mapped soil class, with K = 0.45 t ha h ha^-1 MJ^-1 mm^-1, while Sand was assigned a moderate K-factor value of 0.25 t ha h ha^-1 MJ^-1 mm^-1. This distinction is important because the most widespread soil class is not necessarily the most erodible class. The combination of extensive sandy surfaces and localized higher-erodibility Sandy Loam areas creates a persistent soil-related susceptibility that can amplify runoff-driven erosion during episodic rainfall events. Because soil texture is relatively stable over time, any future increase in storm intensity could exacerbate erosion in these susceptible zones, underscoring the sensitivity of arid soil systems to climatic stressors (Pontes et al., 2022; Preetha and Joseph, 2025).

Topography plays a decisive role in controlling the spatial pattern of soil loss. Most of the watershed is characterized by gentle-to-moderate terrain with relatively low LS-factor values, primarily in the 0–3.82 range. However, localized areas of higher topographic amplification occur where LS values range from 3.83 to 5.79 and 5.80–15.23, especially along steeper western slopes and drainage-related terrain. These areas function as topographic amplifiers of erosion because slope steepness, slope length, and flow convergence increase runoff energy and the capacity for sediment transport. Although large portions of the watershed have limited slope-driven erosion potential, these confined high-LS areas can contribute disproportionately to soil loss. This pattern reinforces the importance of terrain configuration in dryland erosion dynamics and is consistent with studies showing that slope-controlled erosion hotspots often dominate sediment production in otherwise low-gradient catchments (Fentaw and Abegaz, 2024; Ansari and Tayfur, 2023).

Surface cover conditions also strongly influence erosion susceptibility. The C-factor results indicate limited vegetation protection across most of the watershed, with values approaching 1.00 and only scattered areas of lower C-factor values along drainage lines and topographically favorable microsites. Sparse vegetation exposes the soil surface to raindrop impacts, reduces surface roughness, and increases the likelihood of runoff generation, especially where it occurs on steep slopes and over erodible soils. The NDVI and LST time-series analyses support this interpretation by showing persistent low vegetation cover and elevated surface temperatures across much of the watershed. These conditions indicate chronic surface stress rather than short-term disturbance. The SERI results provide supporting information on spatial differences in vegetation deficiency relative to Land Surface Temperature. Higher SERI classes indicate greater vegetation deficiency after thermal scaling and are interpreted as relative surface-susceptibility classes rather than direct measurements of erosion severity or thermal stress alone. This interpretation is consistent with arid-region studies that link low vegetation density, high land surface temperature, and exposed surfaces to land degradation and susceptibility to erosion (Rahimi et al., 2025; Adeyeri et al., 2024; Agnihotri et al., 2021; Musasa et al., 2024; Vatandaşlar and Yavuz, 2022).

The P-factor pattern suggests that formal erosion-control practices are limited across the watershed. Because detailed field data on conservation practices were unavailable, the P-factor was represented using a land-cover and slope-based proxy. High P-factor values across much of the watershed indicate that erosion control depends mainly on natural land-cover conditions rather than engineered conservation measures. This is important because bare and sparsely vegetated surfaces have limited resistance to runoff-driven erosion during intense rainfall events. The spatial overlap between high P-factor values, sparse vegetation, and drainage-linked terrain highlights areas where conservation measures are most needed. Similar conditions have been reported in other arid basins, where limited conservation infrastructure and exposed surfaces increase susceptibility to land degradation (Rashid et al., 2025; Alsaihani and Alharbi, 2024).

The RUSLE-estimated soil-loss results show substantial spatial variability, with annual soil loss ranging from 0 to 319.5 t ha^-1 yr^-1. Slight erosion covers 43.2% of the watershed, indicating that large areas have relatively low modeled soil-loss potential. However, severe erosion exceeding 40 t ha^-1 yr^-1 affects 10.9% of the watershed and is primarily concentrated along drainage corridors, steep slopes, and exposed surfaces in the western and central portions of the study area. This spatial pattern demonstrates that erosion risk is not uniformly distributed. Instead, a limited number of geomorphologically sensitive zones account for the highest modeled erosion potential. These findings align with recent RUSLE-based studies showing that erosion hotspots commonly occur along drainage networks and steep terrain, where localized extremes can dominate geomorphic impacts despite covering relatively small areas (Tarek et al., 2023; Yousuf et al., 2025; Teku and Workie, 2025).

The mapped erosion pattern provides clear implications for watershed management. Conservation strategies should not be applied uniformly across the entire watershed. Instead, interventions should prioritize locations where multiple risk factors overlap, particularly severe erosion zones, high sediment-yield areas, steep western slopes, drainage corridors, and exposed surfaces with limited vegetation cover. In drainage-linked hotspots, check dams, small sediment traps, gully stabilization, and channel-bank protection could reduce runoff energy and sediment transfer. On hillslopes with high modeled soil-loss potential, contour bunding, surface roughness enhancement, runoff-dispersal measures, and localized slope stabilization could help reduce flow concentration and soil detachment. In sparsely vegetated but geomorphologically sensitive areas, revegetation using drought-resistant native species could improve surface protection, increase root reinforcement, and gradually reduce C-factor values. These targeted measures are particularly important under climate-change conditions, where more intense rainfall events could increase runoff energy and accelerate erosion in the mapped hotspots. Because erosion-control measures require long-term planning, maintenance, and institutional coordination, spatial prioritization should also consider the broader socio-hydrological context in which communities, infrastructure, and watershed-management decisions interact (Alruzuq et al., 2023).

These management implications are also relevant to ongoing Saudi and regional environmental initiatives that emphasize vegetation recovery, land restoration, sustainable water-resource management, and rehabilitation of degraded dryland landscapes. In this context, spatially explicit erosion and sediment-delivery maps can help translate broad restoration goals into practical watershed-scale priorities by identifying where conservation measures, revegetation, runoff control, and monitoring should be concentrated (Elkebti and Khalifa, 2025; Ziadat et al., 2025; Hamed et al., 2024; Turki et al., 2023).

The integrated RUSLE–SDR analysis reveals an important difference between soil-loss potential and sediment-delivery potential. Although severe erosion hotspots occur in parts of the watershed, SDR values are generally low, indicating that much of the eroded material is likely retained within the watershed through hillslope deposition, local storage, and channel retention. The SDR range of 0.009–0.151 implies that approximately 85%–99% of eroded material may be retained locally, depending on spatially varying sediment-delivery potential. This erosion-delivery disconnect is characteristic of dryland watersheds and is consistent with broader hydrogeomorphic studies showing that sediment transfer, floodplain connectivity, and channel–floodplain exchange are strongly controlled by hydrologic activation, topography, channel morphology, and sediment storage within the drainage system (Alruzuq and Mossa, 2026; Alruzuq et al., 2026). The finding that less than 1% of the watershed contributes substantially to high sediment-yield zones highlights the importance of distinguishing gross erosion from effective sediment export when designing conservation strategies and downstream sediment management plans (Ghimire et al., 2024; Hou et al., 2024; Yu et al., 2024).

The internal consistency assessment provides additional insight into how the modeled erosion patterns relate to supporting indicators. The weak positive relationship between RUSLE and SERI reflects the fact that SERI represents vegetation deficiency relative to surface temperature rather than directly reproducing RUSLE-estimated soil loss. In contrast, the strong positive relationship between RUSLE and SDR reflects shared topographic controls and the close connection between soil-loss intensity and sediment-delivery potential. The Random Forest analysis yielded an R² of 0.8173, indicating that the RUSLE input factors capture much of the spatial variability in the RUSLE-derived soil-loss output. However, because the RF response variable was generated from the same RUSLE factors used as predictors, this result should be interpreted as an internal model-consistency and sensitivity assessment rather than independent field validation. The feature-importance results, which emphasize LS and the land-cover/slope-based P-factor proxy, indicate that topography and support-practice assumptions exert the strongest influence on modeled soil-loss variability. This interpretation is consistent with RF-based erosion studies in which terrain and land-use factors frequently emerge as important predictors of soil-loss patterns (Tarek et al., 2023; Ge et al., 2023; Sahour et al., 2021).

5.1 Limitations and future research directions

The results should be interpreted in light of several methodological and data-related limitations. The most important limitation is the absence of direct field or gauge-based verification of soil loss, runoff, or sediment yield. Therefore, the RUSLE and SDR outputs should be understood as spatial screening estimates of relative erosion and sediment-delivery potential rather than field-calibrated measurements of actual erosion or outlet sediment export. Future research should incorporate erosion plots, field-based gully inventories, repeated UAV or LiDAR surveys, sediment-load measurements, reservoir or channel sedimentation records, sediment tracing, and ground observations to directly validate modeled erosion and sediment-delivery patterns.

The internal consistency checks also have limitations because they are not independent validation. SERI, SDR, and Random Forest provide useful supporting evidence, but they remain connected to environmental controls already represented in the RUSLE framework. RUSLE and SDR share topographic controls, particularly slope, while the RF model uses RUSLE-derived soil loss as the response variable and RUSLE factors as predictors. These analyses are therefore useful for evaluating model behavior, spatial consistency, and factor sensitivity, but they cannot replace independent field-based validation.

The study focuses on water-induced erosion and does not explicitly quantify wind-driven aeolian erosion. This is an important limitation in hyper-arid environments dominated by dry sandy surfaces, sparse vegetation, and exposed soils. Wind erosion may contribute substantially to land degradation in the study watershed, especially during dry and windy periods. Future studies should integrate water-erosion models, such as RUSLE, with wind-erosion models, such as the Revised Wind Erosion Equation (RWEQ), to provide a more complete assessment of total erosion risk in hyper-arid landscapes. Also, uncertainty is introduced by the scale mismatch among the input datasets. CHIRPS precipitation has a spatial resolution of about 0.05° (approximately 5.5 km), whereas the ALOS PALSAR DEM has a spatial resolution of 12.5 m. This means that rainfall erosivity is represented at a much coarser spatial scale than topography. As a result, the R-factor may smooth localized rainfall variability and may not fully capture small-scale convective storms that strongly influence runoff and erosion in wadi systems. Future work should use local rain-gauge observations, radar rainfall products, or higher-resolution rainfall-intensity datasets when available.

The rainfall erosivity factor was derived from mean annual precipitation rather than storm-event rainfall intensity. This is a limitation because erosion in hyper-arid watersheds is often driven by rare, short-duration, high-intensity storms rather than by average annual rainfall. Without rainfall-intensity data, the R-factor should be interpreted as an annual precipitation-based approximation rather than a direct estimate of event-scale erosive energy. Future research should calculate erosivity using event-based rainfall intensity metrics, such as EI30, where high-temporal-resolution rainfall observations become available.

The K-factor was assigned using texture-based values derived from a 250 m OpenLandMap soil product. This approach was necessary because detailed field measurements of organic matter, soil structure, permeability, aggregate stability, gravel cover, crusting, carbonate content, and particle-size distribution were unavailable. However, texture-only K-factor assignment cannot fully represent local soil variability or surface sealing processes that may strongly affect erodibility in arid environments. Field soil sampling and laboratory measurements would improve K-factor calibration and reduce uncertainty in future applications.

The NDVI-based C-factor also has limitations in bright desert environments. In sparse vegetation settings, NDVI can be strongly affected by soil background brightness, surface crusts, dry vegetation, and mixed pixels. As a result, NDVI may underestimate or overestimate vegetation protection where vegetation cover is very low. Future work should compare NDVI-based C-factor estimates with soil-adjusted vegetation indices, field vegetation cover measurements, and high-resolution imagery to improve estimation of cover-management factors in desert landscapes.

The P-factor was estimated using land-cover and slope-based proxy values because spatially detailed information on actual conservation practices was unavailable. Therefore, the P-factor should be interpreted as a support-practice proxy rather than a mapped inventory of real erosion-control structures. Future research should include field mapping of check dams, terraces, contour bunds, grazing pressure, vehicle tracks, and other land-management practices to improve P-factor representation.

The LULC layer was derived from Google Dynamic World V1 rather than from a locally trained classifier. The 73.8% overall agreement reported by Brown et al. (2022) represents a product-level benchmark for single-image Dynamic World outputs and does not constitute a site-specific accuracy assessment of the temporally aggregated and reclassified LULC map used in this study. Dynamic World may also exhibit confusion among bare ground, shrub and scrub, grass, and cropland in arid environments. Therefore, uncertainty in the reclassified LULC categories may propagate into the P-factor and the resulting RUSLE soil-loss estimates. Future research should validate the derived LULC map using probability-based local reference samples and high-resolution imagery.

Addressing these limitations will strengthen future erosion modeling in hyper-arid watersheds and improve the usefulness of such studies for land restoration, watershed management, and climate-adaptation planning.

6 Conclusion

This study developed an integrated RUSLE–SDR–remote sensing–GeoAI framework to assess water-induced soil loss, sediment-delivery potential, and surface vulnerability in a hyper-arid watershed containing Wadi Samnan near Az Zulfi, Saudi Arabia. The results demonstrate that erosion risk in the study watershed is controlled by the combined effects of episodic rainfall, sandy and locally erodible surface materials, sparse vegetation cover, topographic variability, and limited support-practice conditions. Although much of the watershed is characterized by low to moderate erosion potential, localized hotspots occur where steep slopes, drainage convergence, exposed surfaces, and erodible soils coincide.

The RUSLE results showed that estimated annual soil loss ranged from 0 to 319.5 t ha^-1 yr^-1. Slight erosion covered 43.2% of the watershed, while severe erosion hotspots affected 10.9% of the area and were mainly concentrated along drainage corridors and steeper terrain. These findings indicate that erosion management should be spatially targeted rather than uniformly applied across the entire watershed. Priority should be given to the limited but geomorphologically important zones where soil detachment, runoff concentration, and sediment transfer are most likely to occur.

The SDR and sediment-yield results further showed that high soil-loss potential does not necessarily translate into high sediment export. Most of the watershed exhibited low sediment-delivery ratios, indicating limited sediment connectivity and substantial local retention of eroded material. Approximately 85%–99% of eroded material is likely retained within the watershed through hillslope deposition, local storage, and channel retention, while less than 1% of the watershed contributes substantially to high sediment-yield zones. This distinction between gross soil-loss potential and sediment-delivery potential is particularly important in hyper-arid wadis, where runoff is episodic and sediment movement is strongly controlled by storm intensity, slope conditions, and drainage connectivity.

The SERI comparison and Random Forest analysis provided useful internal consistency and sensitivity evidence, but they should not be interpreted as independent field validation. The Random Forest regression produced an R² of 0.8173, indicating that the RUSLE input factors accounted for 81.73% of the variance in RUSLE-derived soil-loss estimates. Because the RF response variable was derived from the same RUSLE factors used as predictors, this result represents an internal model-consistency and sensitivity assessment rather than validation against observed erosion. Therefore, the study provides a practical spatial screening framework, while future research should incorporate field observations, sediment measurements, high-resolution rainfall data, improved soil measurements, wind-erosion modeling, and independent validation datasets.

Overall, the integration of RUSLE, SDR, remote sensing, GIS, and GeoAI provides a useful framework for identifying erosion-prone areas, distinguishing soil-loss potential from sediment-delivery potential, and supporting sustainable watershed management in data-scarce dryland environments. The mapped erosion and sediment-yield hotspots provide a spatial basis for prioritizing conservation and monitoring efforts, while the identified limitations highlight the key data and methodological improvements needed for future research. In the context of potential climate-driven intensification of extreme rainfall, such spatially explicit erosion assessments can support adaptation planning, land restoration, and targeted erosion-control strategies in hyper-arid watersheds.

References

• 1

  AdeyeriO. E.FolorunshoA. H.AyegbusiI. K.BobdeV.AdeliyiT. E.NdehedeheC. E.et al (2024). Land surface dynamics and meteorological forcings modulate land surface temperature characteristics. Sustain. Cities Soc.101, 105072. 10.1016/j.scs.2023.105072

  • CrossRef
  • Google Scholar
  • View reference in article

• 2

  AdouM.WuY.ZhaoF.QinC. (2024). Drought analysis using normalized difference vegetation index and land surface temperature over Niamey region, the Southwestern of the niger between 2013 and 2019. J. Hydrology Regional Stud.52, 101689. 10.1016/j.ejrh.2024.101689

  • CrossRef
  • Google Scholar
  • View reference in article

• 3

  AgnihotriD.KumarT.JhariyaD. (2021). Intelligent vulnerability prediction of soil erosion hazard in semi-arid and humid region. Environ. Dev. Sustain.23 (2), 2524–2551. 10.1007/s10668-020-00685-2

  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  Al-HashimM. H.Al-AidarosA.ZaidiF. K. (2024). Geological and hydrochemical processes driving karst development in southeastern Riyadh, central Saudi Arabia. Water16 (14), 1937. 10.3390/w16141937

  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  AlqadhiS.MallickJ.TalukdarS.AlkahtaniM. (2023). An artificial intelligence-based assessment of soil erosion probability indices and contributing factors in the abha-khamis watershed, Saudi Arabia. Front. Ecol. Evol.11, 1189184. 10.3389/fevo.2023.1189184

  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  AlruzuqA. R. (2026). GeoAI-Based multi-model framework for spatiotemporal drought modeling in arid environments. Int. J. Remote Sens.0 (0), 1–44. 10.1080/01431161.2026.2689555

  • CrossRef
  • Google Scholar
  • View reference in article

• 7

  AlruzuqA. R.MossaJ. (2025). Machine learning algorithms of riverbed change and environments of the lower apalachicola river. Discov. Appl. Sci.7 (5), 427. 10.1007/s42452-025-06810-y

  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  AlruzuqA. R.MossaJ. (2026). Advancements in large floodplain inundation mapping: integrated computational modeling and remote sensing techniques. Energy Nexus21, 100650. 10.1016/j.nexus.2026.100650

  • CrossRef
  • Google Scholar
  • View reference in article

• 9

  AlruzuqA. R.RemoJ. W. F.MossaJ.AshK. (2023). Socio-hydrology and vulnerability of levee systems along the lower illinois river. Ann. GIS29 (2), 273–291. 10.1080/19475683.2023.2166113

  • CrossRef
  • Google Scholar
  • View reference in article

• 10

  AlruzuqA. R.MossaJ.AmanambuA. C.ChenY.-H.BrennerM. (2026). Morphodynamics and riverbed elevation changes in the lower apalachicola river: a study of large lowland river systems. Acta Geophys.74, 11. 10.1007/s11600-025-01744-w

  • CrossRef
  • Google Scholar
  • View reference in article

• 11

  AlsaihaniM.AlharbiR. (2024). Mapping of soil erosion vulnerability in wadi bin abdullah, Saudi Arabia through RUSLE and remote sensing. Water16 (18), 2663. 10.3390/w16182663

  • CrossRef
  • Google Scholar
  • View reference in article

• 12

  AnandS.KumarH.KumarP.KumarM. (2025). Analyzing landscape changes and their relationship with land surface temperature and vegetation indices using remote sensing and AI techniques. Geosci. Lett.12 (1), 7. 10.1186/s40562-024-003724

  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  AnsariA.TayfurG. (2023). Comparative analysis of estimation of slope-length gradient (LS) factor for entire Afghanistan. Geomatics, Nat. Hazards Risk14 (1), 2200890. 10.1080/19475705.2023.2200890

  • CrossRef
  • Google Scholar
  • View reference in article

• 14

  BahrawiJ. A.ElhagM.AldhebianiA. Y.GalalH. K.HegazyA. K.AlghailaniE. (2016). Soil erosion estimation using remote sensing techniques in wadi yalamlam basin, Saudi Arabia. Adv. Mater. Sci. Eng.2016, 1–8. 10.1155/2016/9585962

  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  BaiddahA.KrimissaS.HajjiS.IsmailiM.AbdelrahmanK.El BouzekraouiM.et al (2023). Head-cut gully erosion susceptibility mapping in semi-arid region using machine learning methods: insight from the high atlas, Morocco. Front. Earth Sci.11, 1184038. 10.3389/feart.2023.1184038

  • CrossRef
  • Google Scholar
  • View reference in article

• 16

  BrownC. F.BrumbyS. P.Guzder-WilliamsB.BirchT.HydeS. B.MazzarielloJ.et al (2022). Dynamic world, near real-time global 10 m land use land cover mapping. Sci. Data9, 251. 10.1038/s41597-022-01307-4

  • CrossRef
  • Google Scholar
  • View reference in article

• 17

  Corral-Pazos-de-ProvensE.Rapp-ArrarásÍ.Domingo-SantosJ. M. (2023). The USLE soil erodibility nomograph revisited. Int. Soil Water Conservation Res.11 (1), 1–13. 10.1016/j.iswcr.2022.07.001

  • CrossRef
  • Google Scholar
  • View reference in article

• 18

  EbabuK.TsunekawaA.HaregeweynN.TsuboM.AdgoE.FentaA. A.et al (2022). Global analysis of cover management and support practice factors that control soil erosion and conservation. Int. Soil Water Conservation Res.10 (2), 161–176. 10.1016/j.iswcr.2021.12.002

  • CrossRef
  • Google Scholar
  • View reference in article

• 19

  EjazN.ElhagM.BahrawiJ.ZhangL.GabrielH. F.RahmanK. U. (2023). Soil erosion modelling and accumulation using RUSLE and remote sensing techniques: case study wadi baysh, Kingdom of Saudi Arabia. Sustainability15 (4), 3218. 10.3390/su15043218

  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  ElkebtiO. A. M.KhalifaW. M. S. (2025). Assessing the Saudi and Middle East green initiatives: the role of environmental governance, renewable energy transition, and innovation in achieving a regional green future. Sustainability17 (12), 5307. 10.3390/su17125307

  • CrossRef
  • Google Scholar
  • View reference in article

• 21

  FentawA. E.AbegazA. (2024). Soil erosion assessment and identification of erosion hotspot areas in the upper tekeze basin, northern Ethiopia. Heliyon10 (12), e32880. 10.1016/j.heliyon.2024.e32880

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  FilchevL.KolevV. (2023). Assessing of soil erosion risk through geoinformation sciences and remote sensing—A review. Earth Environ. Sci. Libr., 377–430. 10.1007/978-3-030-76116-5_21

  • CrossRef
  • Google Scholar
  • View reference in article

• 23

  FirooziA. A.FirooziA. A. (2024). Water erosion processes: mechanisms, impact, and management strategies. Results Eng.24, 103237. 10.1016/j.rineng.2024.103237

  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  GeY.ZhaoL.ChenJ.LiX.LiH.WangZ.et al (2023). Study on soil erosion driving forces by using RUSLE framework and machine learning: a case study in southwest China. Land12 (3), 639. 10.3390/land12030639

  • CrossRef
  • Google Scholar
  • View reference in article

• 25

  GhazanfarS. A. (2024). Biogeography and conservation in the arabian peninsula: a present perspective. Plants13 (15), 2091. 10.3390/plants13152091

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 26

  GhimireS.SinghU.PanthiK. K.BhattaraiP. K. (2024). Spatial sediment erosion and yield using RUSLE coupled with distributed SDR model. Water16 (24), 3549. 10.3390/w16243549

  • CrossRef
  • Google Scholar
  • View reference in article

• 27

  GueffafA.HadjiR.FaqeihK.AlamriS. M.AlameryE.AldubehiM. A.et al (2026). Geospatial assessment and mapping of water-induced soil erosion in a semiarid region of the MENA using GIS-Based RUSLE modeling. Front. Earth Sci.13, 1731125. 10.3389/feart.2025.1731125

  • CrossRef
  • Google Scholar
  • View reference in article

• 28

  HalderJ. C. (2023). The integration of RUSLE-SDR lumped model with remote sensing and GIS for soil loss and sediment yield estimation. Adv. Space Res.71 (11), 4636–4658. 10.1016/j.asr.2023.01.008

  • CrossRef
  • Google Scholar
  • View reference in article

• 29

  HamedY.AyadiY.KhalilR.Al-OmranA.LebdiF.DhaouadiL. (2024). Wastewater resources, agricultural practices management strategies, soil salinity predictions and artificial recharge in the middle east-saudi arabia: a review. J. Saudi Soc. Agric. Sci.23 (8), 569–584. 10.1016/j.jssas.2024.08.003

  • CrossRef
  • Google Scholar
  • View reference in article

• 30

  HassaniA.AzapagicA.ShokriN. (2020). Predicting long-term dynamics of soil salinity and sodicity on a global scale. Proc. Natl. Acad. Sci.117 (52), 33017–33027. 10.1073/pnas.2013771117

  • CrossRef
  • Google Scholar
  • View reference in article

• 31

  HouX.YangH.CaoJ. (2024). The spatial and temporal dynamics of soil conservation and its influencing factors in the ten tributaries of the upper yellow river, China. Water16 (20), 2888. 10.3390/w16202888

  • CrossRef
  • Google Scholar
  • View reference in article

• 32

  KarimiH.SultanM.YanE.ElhaddadH.SalehH.AbdelmohsenK.et al (2025). Climate-extreme modeling framework for sustainable flood management in the arabian peninsula. J. Environ. Manag.393, 127074. 10.1016/j.jenvman.2025.127074

  • CrossRef
  • Google Scholar
  • View reference in article

• 33

  KimH. S.JulienP. Y. (2006). Soil erosion modeling using RUSLE and GIS on the IMHA watershed. Water Eng. Res.7 (1), 29–41.

  • Google Scholar
  • View reference in article

• 34

  KousarS.ShiraziS. A. (2023). Spatio-temporal variations of land-use land-cover in response to RUSLE model in swat basin using remote sensing and GIS techniques. Sarhad J. Agric.39 (3). 10.17582/journal.sja/2023/39.3.722.737

  • CrossRef
  • Google Scholar
  • View reference in article

• 35

  MooreI. D.BurchG. J. (1986). Modelling erosion and deposition: topographic effects. Trans. ASAE29 (6), 1624–1630. 10.13031/2013.30363

  • CrossRef
  • Google Scholar
  • View reference in article

• 36

  MusasaT.DubeT.MarambanyikaT. (2024). Landsat satellite programme potential for soil erosion assessment and monitoring in arid environments: a review of applications and challenges. Int. Soil Water Conservation Res.12 (2), 267–278. 10.1016/j.iswcr.2023.10.003

  • CrossRef
  • Google Scholar
  • View reference in article

• 37

  OdnoletkovaN.PatzekT. W. (2023). Water resources in Saudi Arabia: trends in rainfall, water consumption, and analysis of agricultural water footprint. Npj Sustain. Agric.1 (1), 7. 10.1038/s44264-023-00006-w

  • CrossRef
  • Google Scholar
  • View reference in article

• 38

  PanagosP.BorrelliP.MeusburgerK.ZandenE. H.PoesenJ.AlewellC. (2015). Modelling the effect of support practices (P-factor) on the reduction of soil erosion by water at European scale. Environ. Sci. & Policy51, 23–34. 10.1016/j.envsci.2015.03.012

  • CrossRef
  • Google Scholar
  • View reference in article

• 39

  PanagosP.BorrelliP.MatthewsF.LiakosL.BezakN.DiodatoN.et al (2022). Global rainfall erosivity projections for 2050 and 2070. J. Hydrology610, 127865. 10.1016/j.jhydrol.2022.127865

  • CrossRef
  • Google Scholar
  • View reference in article

• 40

  PatlakasP.StathopoulosC.FlocasH.BartsotasN. S.KallosG. (2021). Precipitation climatology for the arid region of the arabian peninsula—variability, trends and extremes. Climate9 (7), 103. 10.3390/cli9070103

  • CrossRef
  • Google Scholar
  • View reference in article

• 41

  PontesS. F.JacquesY.MartinsV.BoechatC. L.FerreiraS.DantasJ. S.et al (2022). Prediction of soil erodibility by diffuse reflectance spectroscopy in a neotropical dry forest biome. Land11 (12), 2188. 10.3390/land11122188

  • CrossRef
  • Google Scholar
  • View reference in article

• 42

  PreethaP.JosephN. (2025). Evaluating modified soil erodibility factors with the aid of pedotransfer functions and dynamic remote-sensing data for soil health management. Land14 (3), 657. 10.3390/land14030657

  • CrossRef
  • Google Scholar
  • View reference in article

• 43

  RahamanM.SouthworthJ.AmanambuA. C.TeferaB. B.AlruzuqA. R.SafaeiM.et al (2025). Combining deep learning and machine learning techniques to track air pollution in relation to vegetation cover utilizing remotely sensed data. J. Environ. Manag.376, 124323. 10.1016/j.jenvman.2025.124323

  • CrossRef
  • Google Scholar
  • View reference in article

• 44

  RahimiE.DongP.JungC. (2025). Global NDVI-LST correlation: Temporal and spatial patterns from 2000 to 2024. Environments12 (2), 67. 10.3390/environments12020067

  • CrossRef
  • Google Scholar
  • View reference in article

• 45

  RashidM.HaiderS.RizwanA.NaseerM. W.AslamM. F.HamzaM.et al (2025). Mitigating soil erosion in arid landscapes: integrating RUSLE and geospatial analysis for sustainable land management. Environ. Challenges20, 101210. 10.1016/j.envc.2025.101210

  • CrossRef
  • Google Scholar
  • View reference in article

• 46

  RashmiI.KarthikaK. S.RoyT.ShinojiK. C.KumawatA.KalaS.et al (2022). Soil erosion and sediments: a source of contamination and impact on agriculture productivity. Agrochem. Soil Environ., 313–345. 10.1007/978-981-16-9310-6_14

  • CrossRef
  • Google Scholar
  • View reference in article

• 47

  RenardK. G.FosterG. R.WeesiesG. A.McCoolD. K.YoderD. C. (1997). Predicting Soil Erosion by Water: A Guide to Conservation Planning with the Revised Universal Soil Loss Equation (RUSLE). Washington, DC: U.S. Department of Agriculture. Agriculture Handbook No. 703.

  • Google Scholar
  • View reference in article

• 48

  SahourH.GholamiV.VazifedanM.SaeediS. (2021). Machine learning applications for water-induced soil erosion modeling and mapping. Soil Tillage Res.211, 105032. 10.1016/j.still.2021.105032

  • CrossRef
  • Google Scholar
  • View reference in article

• 49

  SahuH.SinghR.KumarU.AlruzuqA.PandeC. B. (2024). “A comparative assessment of organic and inorganic farming impact on land surface temperature from 1991 to 2021 in the decade of Punjab and Uttarakhand,” in Natural Resource Monitoring, Planning and Management Based on Advanced Programming. Editors MishraA. P.KaushikA.PandeC. B. (Singapore: Springer Nature), 249–267. 10.1007/978-981-97-2879-4_13

  • CrossRef
  • Google Scholar
  • View reference in article

• 50

  SahuH.NagarkotiJ.SardarP.PurohitP.MishraA. P.SarkarM. S.et al (2025). “Evaluating soil erosion in dehradun using the RUSLE model: challenges, impacts, and policy strategies for effective soil conservation,” in Remote Sensing for Environmental Monitoring. Editors SarithaV.PandeC. B.SinghR.ShahidM. (Singapore: Springer Nature), 239–261. 10.1007/978-981-96-5546-5_12

  • CrossRef
  • Google Scholar
  • View reference in article

• 51

  SchmidtS.TreschS.MeusburgerK. (2019). Modification of the RUSLE slope length and steepness factor (LS-factor) based on rainfall experiments at steep alpine grasslands. MethodsX6, 219–229. 10.1016/j.mex.2019.01.004

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 52

  SenS.AlmusabehA.AbouelreshM. O. (2023). Geoheritage and geotourism potential of tuwaiq Mountain, Saudi Arabia. Geoheritage15 (3), 93. 10.1007/s12371-023-00861-6

  • CrossRef
  • Google Scholar
  • View reference in article

• 53

  SouthworthJ.SmithA. C.SafaeiM.RahamanM.AlruzuqA.TeferaB. B.et al (2024). Machine learning versus deep learning in land system science: a decision-making framework for effective land classification. Front. Remote Sens.5, 1374862. 10.3389/frsen.2024.1374862

  • CrossRef
  • Google Scholar
  • View reference in article

• 54

  TarekZ.ElsheweyA. M.ShohiebS. M.ElhadyA. M.El-AttarN. E.ElseuofiS.et al (2023). Soil erosion status prediction using a novel random forest model optimized by random search method. Sustainability15 (9), 7114. 10.3390/su15097114

  • CrossRef
  • Google Scholar
  • View reference in article

• 55

  TekuD.WorkieM. D. (2025). Longitudinal analysis of soil erosion dynamics using the RUSLE model in Ethiopia’s Lake ziway watershed: implications for agricultural sustainability and food security. Front. Environ. Sci.12, 1506001. 10.3389/fenvs.2024.1506001

  • CrossRef
  • Google Scholar
  • View reference in article

• 56

  ThapaP. (2020). Spatial estimation of soil erosion using RUSLE modeling: a case study of dolakha district, Nepal. Environ. Syst. Res.9 (1), 15. 10.1186/s40068-020-00177-2

  • CrossRef
  • Google Scholar
  • View reference in article

• 57

  TurkiF.MasmoudiF.KaririE. (2023). “Data science approach for climate change in Saudi Arabia: trend analysis,” in 2023 Sixth International Conference of Women in Data Science at Prince Sultan University (Wids PSU), 1–6. 10.1109/WiDS-PSU57071.2023.00014

  • CrossRef
  • Google Scholar
  • View reference in article

• 58

  VanoniV. A. (1975). Sedimentation Engineering Practice. American Society of Civil Engineers.

  • Google Scholar
  • View reference in article

• 59

  VatandaşlarC.YavuzM. (2022). Useful indicators and models for assessing erosion control ecosystem service in a semi-arid forest landscape. Environ. Monit. Assess.195 (1), 208. 10.1007/s10661-022-10814-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 60

  WallingD. E. (2017). “Measuring sediment yield from river basins,” in Soil Erosion Research Methods (London, UK: Routledge), 39–82. Available online at: https://www.taylorfrancis.com/chapters/edit/10.1201/9780203739358-3/measuring-sediment-yield-river-basins-walling (Accessed August 7, 2026).

  • Google Scholar
  • View reference in article

• 61

  YangX.YoungJ.ShiH.ZhuQ.PulsfordI.ChapmanG.et al (2024). Estimating sediment delivery ratio using the RUSLE-IC-SDR approach at a complex landscape: a case study at the lower snowy river area, Australia. J. Hydrology645, 132237. 10.1016/j.jhydrol.2024.132237

  • CrossRef
  • Google Scholar
  • View reference in article

• 62

  YaswanthK.KonaM.AndraS. K.RathinasamyM. (2022). Understanding the impact of changes in land-use land-cover and rainfall patterns on soil erosion rates using the RUSLE model and GIS techniques: a study on the nagavali river basin. J. Water Clim. Change13 (7), 2648–2670. 10.2166/wcc.2022.016

  • CrossRef
  • Google Scholar
  • View reference in article

• 63

  YousufA.SharmaN.BhatM. A.KhokharA.SandhuP. S.DhillonB. S.et al (2025). Soil erosion assessment and watershed prioritization using GIS-Based RUSLE in lower shivaliks of northwest India. Geol. Ecol. Landscapes10 (1), 294–311. 10.1080/24749508.2025.2450108

  • CrossRef
  • Google Scholar
  • View reference in article

• 64

  YuJ.ZhaoQ.YuZ.LiuY.DingS. (2024). A review of the sediment production and transport processes of forest road erosion. Forests15 (3), 454. 10.3390/f15030454

  • CrossRef
  • Google Scholar
  • View reference in article

• 65

  ZiadatF.ConcheddaG.HaddadF.NjeruJ.BrèsA.DawelbaitM.et al (2025). Desertification and agrifood systems: restoration of degraded agricultural lands in the Arab region. Agriculture15 (12), 1249. 10.3390/agriculture15121249

  • CrossRef
  • Google Scholar
  • View reference in article