Mapping Soil Erosion and Ecosystem Service Loss: Integrating RUSLE and NDVI Metrics to Support Conservation in El Cajas National Park, Ecuador

Abstract

Mountain protected areas in the tropical Andes experience localized yet severe soil erosion that threatens erosion-regulating services and downstream water–energy security. We mapped soil loss at 30 m using the Revised Universal Soil Loss Equation (RUSLE) and quantified the erosion-control service in El Cajas National Park, Ecuador (28,544 ha) using an NDVI-based index. Replacing categorical land cover C factors with a continuous NDVI surface increased the park-wide soil loss estimate by ∼58%, yielding an area-weighted mean of 5.3 t ha⁻¹ yr⁻¹ and local maxima of 120 t ha⁻¹ yr⁻¹ on steep and sparsely vegetated escarpments. Relative to a bare soil scenario, existing páramo grasslands, shrub mosaics, and scattered Polylepis woodlots avert 95% of potential erosion, quantifying the service supplied by vegetation. Between 2023 and 2024, a ∼60% rise in mean NDVI more than doubled the area delivering moderate-to-high erosion control. A hot-spot analysis further identified ∼30 km² (≈5% of the park) where high modeled soil loss coincides with low protection; these clusters generate ∼80% of predicted sediment and constitute priority targets for restoration or visitor use regulation. The integrated RUSLE–NDVI–EC approach provides a concise and transferable screening tool for aligning conservation investments with Ecuador’s restoration pledges and for safeguarding critical hydrological services in Andean protected areas.

1. Introduction

Soil erosion is a critical global environmental problem, leading to the loss of fertile topsoil, reduced agricultural productivity, and damage to ecosystems and infrastructure [1,2]. An estimated 24 billion tons of fertile soil are eroded from the land each year, undermining food security and costing the world economy billions [3]. Indeed, recent global assessments estimate that soil erosion by water imposes an annual cost of around USD 8 billion in lost global GDP and removes over 33 million tonnes of food from production [4]. In addition to these economic losses, erosion-driven sedimentation can clog waterways and reservoirs and disrupt aquatic habitats, while the loss of soil nutrients and organic carbon diminishes land productivity and contributes to land degradation. Collectively, these impacts highlight soil erosion as both an economic and ecological threat on a worldwide scale [5,6].

Mountainous regions are especially prone to accelerated soil erosion, even in areas designated for conservation. High mountain environments such as the Andes experience some of the highest erosion rates on the planet due to their steep slopes, erodible soils, and intense climatic events [7,8]. In tropical Andean landscapes, heavy rainfall on vulnerable slopes can rapidly mobilize soil, particularly where vegetation is sparse or disturbed. Even protected mountain areas are not immune to these processes, as localized human activities such as road building, grazing, or tourist trekking can trigger severe erosion on steep terrain [9,10]. For example, field experiments in the Ecuadorian Andes have shown that significant soil loss on hiking trails can begin after just minutes of intense rainfall, with steeper unvegetated paths experiencing the most rapid erosion [11]. These findings highlight that even conserved mountain areas are at risk of erosion if soils and vegetation are not properly managed.

The Revised Universal Soil Loss Equation (RUSLE) is an empirical model that estimates long-term average soil erosion based on rainfall erosivity, soil erodibility, slope steepness/length, land cover, and conservation practices [12]. Widely applied in highland regions, RUSLE maps erosion risk and identifies hotspots. In the Andes, RUSLE-based modeling in a deforested watershed revealed dramatic increases in soil loss compared to naturally vegetated areas [13]. Converting steep montane forest slopes to cropland has been shown to raise annual soil loss from negligible levels to extreme rates exceeding 900 t/ha/yr on exposed slopes [14,15,16]. RUSLE’s utility in mountain contexts is evident when evaluating how land use changes amplify erosion. For classification of these rates by slope category, we followed FAO (2006) and USDA NRCS (2017) [17,18]. This capacity to rank erosion severity underpins RUSLE’s utility in evaluating how land use changes amplify soil loss in high-relief landscapes.

Vegetation cover is critical for soil erosion control, functioning as a regulating ecosystem service. Plants intercept raindrops and reinforce soil stability with their roots, thereby reducing runoff and sediment yield. Andean watersheds with intact forests or páramo maintain low erosion rates even under conditions involving intense rainfall and steep slopes, whereas deforested areas suffer markedly higher losses [19,20]. Thus, forests and alpine grasslands act as natural barriers to sediment transport, preserving on-site soil fertility and sustaining downstream services such as water purification and reservoir longevity [21,22].

However, traditional planning approaches have often treated soil erosion risk mapping and ecosystem service valuation as separate exercises, potentially missing opportunities to combine insights from both. Many ecosystem service assessments rely on coarse land use proxies to infer services such as erosion control [21,23]. This can oversimplify reality, as the actual service delivered (soil retained) depends on complex biophysical processes and where erosion would occur in the absence of protective cover. Recent studies have started to bridge this divide by integrating RUSLE with cloud-based remote sensing workflows (e.g., Google Earth Engine) and dynamic vegetation indices in order to refine the C factor and map erosion risk at scale [24,25,26]. In parallel, ecosystem service models such as InVEST’s sediment retention explicitly estimate avoided sediment export and link biophysical erosion to service provision [27].

There is a growing recognition that integrating biophysical erosion models with ecosystem service frameworks can yield a more nuanced understanding for decision-making [23]. Integrating RUSLE, which quantifies soil loss under varying conditions, with ecosystem service assessments that translate retained soil into tangible benefits (e.g., sustained crop productivity, enhanced water quality, and flood mitigation) pinpoints areas where conservation yields the greatest return [19]. Yet, applications in tropical Andean protected areas that couple (i) a dynamic NDVI-derived C surface with (ii) an explicit erosion control (proportion of avoided loss) metric and (iii) a hot-spot overlay to guide management remain scarce; this is the specific gap that our study addresses.

El Cajas National Park (CNP) (3000–4450 m above sea level) in southern Ecuador combines páramo grasslands and cloud forests that underpin regional water security by regulating river flow and stabilizing steep terrain [28,29]. Despite its protected status, accelerated erosion and vegetation loss driven by grazing, páramo burning, logging, and limited agriculture have triggered localized landslides, sedimentation, and degraded water quality [30]. Research quantifying soil loss rates and the mitigating role of ecosystem services within CNP remains scarce. The erosion control (EC) index synthesizes key ecohydrological processes by combining mean canopy greenness ($\overline{\mathrm{NDVI}}$), vegetation stability, and terrain susceptibility, thereby quantifying the proportion of potential soil loss intercepted by vegetation. This approach has been successfully applied to map erosion-regulating services in Andean páramo and other steep landscapes [31,32,33].

This study aims to couple RUSLE-based soil loss estimates with ecosystem service indicators to support conservation planning in CNP. We model the spatial distribution of potential soil erosion under current land cover and land use change scenarios. We then integrate these results with an ecosystem service assessment of erosion regulation, identifying where natural vegetation prevents soil loss, providing related benefits such as reduced sediment yield in water bodies. By combining the biophysical model outputs with an ecosystem service framework, we highlight priority areas for erosion control services. This integrated analysis can provide park managers and regional planners with science-based insights to inform strategies that enhance both ecosystem health and human wellbeing. In this way, our research addresses a local management need while serving as a model for applying erosion risk modeling and ecosystem service values for sustainable conservation in mountainous protected areas.

2. Materials and Methods

2.1. Study Area

CNP occupies 285.4 km² (28,544 ha) of the western cordillera of the Andes, about 35 km west of the city of Cuenca in Azuay Province, southern Ecuador. The protected area occupies a rectangular extent in UTM Zone 17S coordinates, ranging from Easting 739,000 m to 759,000 m and Northing 6,867,000 m to 6,888,500 m. It spans a rugged altitudinal gradient from 3138.8 m to 4451.68 m a.s.l. (Figure 1). This topographic amplitude creates a mosaic of high-Andean landforms, U-shaped glacial valleys, rock outcrops, and rolling páramo plateaus that strongly condition erosion processes and ecosystem functioning [28].

The regional climate is cool–humid and markedly seasonal. Long-term data from seven Ecuadorian National Meteorological and Hydrological service (INAMHI) stations show mean annual precipitation between 881 mm and 1466 mm, with a pronounced wet peak in April and a dry minimum in August [34]. Although average air temperatures were not monitored inside the park, páramo stations at similar elevations record large diurnal ranges, with sub-freezing nights and occasional midday maxima above 25 °C. Hydrologically, more than 230 glacial and tectonic lakes act as natural reservoirs that give rise to the Tomebamba, Yanuncay, Mazán, Migüir, Paute, Balao and Cañar catchments, supplying drinking water to Cuenca and driving the Paute hydroelectric complex.

Vegetation cover is typical of equatorial high-Andean ecosystems. Treeless páramo grasslands (Calamagrostis, Festuca, Stipa spp.) blanket 84% of the surface, interspersed with Andean shrubland (8%), Polylepis/Gynoxys woodlots (2%), a dense lacustrine network (3%), wetlands and moss-covered turfs (less than 2%), and only negligible cropland or urban footprints. Such cover regulates infiltration, carbon sequestration and micro-climate, but exposes highly erodible soils and steep slopes to intense rainfall in places where cover is sparse or disturbed [35].

CNP has been part of Ecuador’s national protected-area system since 1977, and its 2018 management-plan update stresses mitigation of soil erosion as a priority to secure downstream water quality and hydropower reliability. Conventional agriculture and engineering terraces are prohibited; accordingly, the RUSLE P-factor was fixed at 1, reflecting an absence of artificial conservation practices. Therefore, the combination of steep relief, high rainfall erosivity, and strategic hydrological services makes erosion control an essential ecosystem service and management objective in El Cajas [36].

2.2. Datasets

2.2.1. Topography and Soils

A 30 m–resolution Digital Elevation Model (DEM) for CNP was downloaded from the national SIGTIERRAS portal (available at http://www.sigtierras.gob.ec accessed on 2 February 2025), curated by Ecuador’s Ministry of Agriculture and Livestock (MAG). The layer served as the basis for slope length ($LS$) calculations and all subsequent raster harmonisation steps. DEM voids over water bodies were filled with bilinear interpolation; final outputs were resampled to a common 30 m grid for integration with other factors [37,38].

Soil information was extracted from the national soil taxonomy map published by MAG, which delineates five dominant orders inside the park: inceptisols (82%), entisols (13%), alfisols (4%), histosols (1%), and minor rock outcrops [39]. For each soil polygon, laboratory-derived texture classes (sand, silt, clay percentages) and organic carbon contents reported in the theses were linked to the map attributes. These values fed directly into the Wischmeier–Smith erodibility equation, yielding spatially variable K factors that ranged from 0.001 for bare rock to 0.055 t h MJ⁻¹ mm⁻¹ ha⁻¹ for organic-rich histosols. All soil polygons were rasterised at 30 m to match the DEM, ensuring pixel-level consistency across the modeling workflow.

2.2.2. Climate

Daily precipitation records (1984–2023) were obtained from the Prediction of Worldwide Energy Resources (POWER) ProjectNASA Langley Research Center, Version 2.5.14; these data were obtained from the NASA Langley Research Center (LaRC) POWER Project funded through the NASA Earth Science/Applied Science Program (available at https://power.larc.nasa.gov).

For each 0.5° grid cell overlapping CNP, the variable PRECTOTCORR (bias-corrected precipitation, mm day⁻¹) was downloaded and aggregated to monthly totals. To reduce the coarse-grid bias over complex topography, the gridded series were linearly bias-adjusted against the seven long-term INAMHI rain gauges (Surucucho, Piscícola Chirimichay, Sayausí, Quinoas, Ricaurte, Chaucha and Buenos Aires). Bias factors (ratio of observed to POWER precipitation) were calculated for each month and applied to the corresponding POWER cell; the bias-corrected values were then distributed across the park by ordinary kriging with a spherical variogram and a 10 km search radius, yielding a 1 km precipitation raster stack.

Rainfall erosivity (R) was computed pixel-wise with the Modified Fournier Index (MFI) following Wischmeier and Smith (1978) (Equations (1) and (2)) [40]:

$$\mathrm{MFI}=\sum _{m=1}^{12}{\displaystyle \frac{P_m^2}{P_{\mathrm{a}}}},$$

(1)

$$R=95\mathrm{MFI}-150,$$

(2)

where $P_m$ is monthly precipitation (mm) and $P_{\mathrm{a}}$ the annual total. The resulting 1 km R surface (MJ mm ha⁻¹ h⁻¹ yr⁻¹) was finally downscaled to the 30 m modelling grid by bilinear resampling, ensuring spatial consistency with the other RUSLE factors.

2.2.3. Satellite Imagery

Normalised Difference Vegetation Index (NDVI) and land cover inputs were derived entirely within the Google Earth Engine (GEE) cloud platform [41]. Monthly NDVI mosaics for 2023 and 2024 were generated from Landsat 8 (LC08) and Landsat 9 (LC09) Level 2, Collection 2, Tier 1 surface-reflectance scenes (path/row 010/062) [42]. Scenes were filtered by acquisition month, atmospherically corrected, and cloud-masked with the QA_PIXEL bit-mask; the median composite of each monthly image collection was then converted to NDVI using the SR_B5 (NIR) and SR_B4 (Red) bands. A total of 22 cloud-filtered monthly rasters (11 per year) were exported at 30 m resolution for subsequent processing. Descriptive statistics of NDVI revealed values from 0.00 to 0.33 in 2023 and 0.01 to 0.29 in 2024 with standard deviations of 0.02 in both years, indicating moderate seasonal variability of vegetation greenness across the park.

Land cover information was obtained from the ESA WorldCover 10 m v100 product (epoch 2020–2021). The eleven WorldCover classes present in CNP, dominated by grassland (84.1%), shrubland (8.4%), open water (3.3%) and high-Andean tree cover (2.2%), were reclassified to the five vegetation categories required for the RUSLE C-factor (tree, shrub, grass, bare, water) and resampled to the 30 m modeling grid. To check positional accuracy, a subset of fifty random points was visually compared against cloud-free Sentinel-2 MSI imagery (10 m) from August 2023; this was a small internal visual cross-check intended as a qualitative sanity check of the five reclassified categories (tree/shrub/grass/bare/water). For product characteristics and independent accuracy reporting, we refer to the ESA WorldCover documentation [43].

We selected Landsat 8–9 surface-reflectance data for 2023–2024 in order to capture the most recent vegetation dynamics and disturbance–recovery patterns at 30 m resolution, leveraging GEE’s global atmospherically corrected mosaics. In contrast, the ESA WorldCover 10 m v100 layer (2020–2021) was used for its higher spatial detail and validated consistency in broad land cover classes; structural categories within this protected area are unlikely to have shifted significantly between 2020 and 2024. Although this introduces a temporal offset, our core modeling relies on the dynamic, year-specific NDVI-derived C-factor, while the static WorldCover map only provides baseline categorical C values. Sensitivity tests comparing static versus NDVI-derived C surfaces indicated that the temporal mismatch affects overall soil loss estimates by less than 5%, demonstrating minimal bias in our results.

All preprocessing, scene filtering, cloud masking, NDVI computation, mosaicking, and WorldCover reclassification were carried out in Google Earth Engine with custom JavaScript routines.

2.3. RUSLE Modeling Workflow

Soil loss for each 30 m cell was estimated with the Revised Universal Soil Loss Equation (Equation (3)):

$$A=R\times K\times LS\times C\times P$$

(3)

where R and K correspond to the rainfall run-off erosivity and soil erodibility surfaces already described in Section 2.2.1 and Section 2.2.2. Both rasters were projected to EPSG 32717 and resampled to the common 30 m grid.

The topographic factor ($LS$) was derived from the 30 m DEM according to Equation (4):

$$LS={\left({\displaystyle \frac{\lambda }{22.13}}\right)}^m\left(0.065+4.56sin\alpha +65.41{sin}^2\alpha \right)$$

(4)

where $\lambda$ is slope length (m), $\alpha$ is slope angle (°), and the exponent m varies from 0.2 on gentle slopes to 0.5 on slopes steeper than 5%.

To establish a baseline estimate ($A_{\mathrm{current}}$), we first applied standard C values from literature to categorical land cover classes [40]. Subsequently, we refined the model by replacing the static C surface with a dynamic NDVI-derived C factor raster (Section 2.4), which better captures subpixel vegetation variability and recent changes in land cover. While these two methods differ in resolution and temporal sensitivity, their use in parallel allows us to assess relative differences in erosion risk under varying vegetation detail rather than producing directly comparable absolute values.

Because engineering terraces and similar conservation measures are prohibited inside the national park, the conservation-practice coefficient was set uniformly to $P=1.0$.

All five factor rasters (R, K, $LS$, C, P) were stacked and multiplied according to Equation (3) in QGIS 3.36, producing the baseline soil loss surface $A_{\mathrm{current}}$ at 30 m resolution.

2.4. Erosion Control Index (EC)

Vegetation’s capacity to buffer soil loss was quantified using an NDVI–based erosion control indicator adapted from the framework proposed by Carreno et al. [31], Hmimina et al. [44], and Pettorelli [32]. Following these authors, the index combines the mean greenness of the vegetation, its temporal stability, and the local steepness of the terrain (Equation (5)):

$$EC=\overline{\mathrm{NDVI}}\left(1-c{v}_{\mathrm{NDVI}}\right)\left(1-P_d\right)$$

(5)

where $\overline{\mathrm{NDVI}}$ is the annual mean of the monthly NDVI, $c{v}_{\mathrm{NDVI}}$ is the coefficient of variation of NDVI through the year (a proxy for canopy stability), and $P_d$ is the dimensionless slope factor obtained from the DEM and is min–max scaled to the interval $[0,1]$.

Cloud-masked surface reflectance scenes (one per month when available, January–December 2023 and 2024) were selected from Landsat 8 and Landsat 9 Tier 1 collections in Google Earth Engine. Each scene was atmospherically corrected with the Landsat C2L2 algorithm and cleared of clouds and cirrus using the QA_PIXEL bit-mask. Monthly mosaics were produced with a median reducer at 30 m resolution and NDVI was computed as Equation (6):

$$\mathrm{NDVI}={\displaystyle \frac{\mathrm{NIR}-\mathrm{RED}}{\mathrm{NIR}+\mathrm{RED}}}$$

(6)

where NIR corresponds to band 5 and RED to band 4 of the Landsat Operational Land sensor. The resulting twelve-layer stack was exported as 16-bit GeoTIFF and re-projected to UTM 17S in order to coincide with the RUSLE grid.

The raw NDVI values (standard range –1 to +1) were then linearly rescaled to a 0–100 range using QGIS’s native “Rescale values” raster operation, facilitating the interpretation, visualization, and modeling framework ([21,22,45]).

Slope in degrees was derived from the 30 m DEM described in Section 2.2.1 using the Horn algorithm. To make slope dimensionless and comparable with NDVI, values were normalized with a linear transformation setting 0° = 0 and the 99th percentile (46.2°) = 1, yielding the term $P_d$ in Equation (5).

For every pixel, the monthly mean NDVI ($\overline{\mathrm{NDVI}}$) and its coefficient of variation ($c{v}_{\mathrm{NDVI}}$) were calculated in R using the terra package [46]. Equation (5) was then applied raster-wise to obtain a continuous EC surface ranging from less than 0.1 (steep, sparsely vegetated scree) to greater than 0.8 (dense, stable páramo tussock on gentle slopes). The EC raster was finally resampled to exactly match the 30 m grid of the updated RUSLE outputs, facilitating subsequent hot-/cold-spot and scenario analyses.

2.5. Integration Framework

The six-step workflow used to couple soil loss with erosion-control service is summarized in

Table 1. Practical integration steps for coupling soil loss and erosion control services.

Step	Action	Expected Product
1	Harmonise grids (projection, 30 m pixel) for all layers	Common raster stack
2	Re-run RUSLE with updated C	Updated soil-loss map $A_{\mathrm{updated}}$
3	Create no-vegetation scenario ($C=1$, $P=1$) $\to A_{\max }$	Maximum soil-loss surface
4	Derive erosion-control index $EC=(A_{\max }-A_{\mathrm{updated}})/A_{\max }$	EC raster (0–1)
5	Hot/cold-spot (Getis-Ord $G_i^{*}$) on high $A_{\mathrm{updated}}$ + low $EC$	Priority map, statistics
6	Scenario analysis: NDVI/LULC shifts	Policy-relevant scenarios

. Briefly, all rasters (R, K, $LS$, initial C, P, monthly NDVI, EC, and LULC) were first snapped to a common 30 m UTM 17S grid, continuous layers via bilinear resampling, and categorical layers via nearest neighbor. Replacing the static C layer with the NDVI-based raster yielded an updated soil loss map ($A_{\mathrm{updated}}$). A “no-vegetation” run ($C=1$, $P=1$) generated the worst-case surface ($A_{\max }$), from which the pixel-level erosion control index was computed as Equation (7):

$$EC={\displaystyle \frac{A_{\max }-A_{\mathrm{updated}}}{A_{\max }}}.$$

(7)

Local clustering of high $A_{\mathrm{updated}}$ + low $EC$ was detected with the Getis–Ord $G_i^{*}$ statistic (queen contiguity, eight neighbors). Two illustrative land-management scenarios (a 10% shrub expansion on slopes 25% and a 15% NDVI reduction in grassland) were propagated through the workflow by recalculating C, re-running RUSLE, and re-estimating $EC$. All raster algebra was performed in QGIS 3.36; spatial statistics and scenario scripts were executed in R 4.3.3 (terra, spdep, sf) [46,47].

3. Results

3.1. Baseline Soil Loss ($A_{\mathrm{current}}$)

The baseline RUSLE map (Figure 2) exhibits a strongly right-skewed distribution of soil loss, with pixel values ranging from near 0 t ha⁻¹ yr⁻¹ (open water) up to 285 t ha⁻¹ yr⁻¹ on the steepest sparsely vegetated scarps. The area-weighted mean and median are 5.3 t ha⁻¹ yr⁻¹ and 0.9 t ha⁻¹ yr⁻¹, respectively, confirming that the most severe erosion is confined to a small portion of the park.

Vegetation classes show that grass-dominated páramo (<5 t ha⁻¹ yr⁻¹), shrubland (4–6 t ha⁻¹ yr⁻¹), and sparse cover (>60 t ha⁻¹ yr⁻¹) correspond to map areas of 84%, 8.4%, and <2%, respectively. Detailed statistics by vegetation and slope are consolidated in

Table 2. Baseline soil loss by slope category. Slope classes are defined following FAO (2006) and USDA NRCS (2017).

Slope Category	Area Fraction (%)	Mean ${\mathit{A}}_{\mathbf{current}}$ (t ha−1 yr−1)
<5% (flat–gentle)	7.0	0.6
5–10% (moderate)	25.6	3.4
10–25% (strong)	41.8	12.1
25–50% (very strong)	21.3	38.4
>50% (escarpment)	4.4	119.7

.

Slope classes follow established guidelines [17,18], grouping gradients into five ranges relevant to erosion processes: negligible (less than 5%), moderate (5–10%), strong (10–25%), very strong (25–50%), and escarpment (>50%). Table 2 summarizes $A_{\mathrm{current}}$ by slope class.

Aggregating raster outputs to the Tomebamba and Yanuncay outlet stations yields a Nash–Sutcliffe efficiency of 0.63 and an RMSE of 18 t yr⁻¹ for 2010–2021, relative to INAMHI sediment gauge records. Because RUSLE produces long-term annual soil loss estimates, model skill is assessed on annual totals; therefore, we do not diagnose seasonal biases.

Present vegetation already prevents roughly four-fifths of potential soil loss. Consequently, erosion control efforts should focus on the narrow belt of steep sparsely vegetated slopes, where even modest restoration could substantially reduce sediment delivery downstream. Present vegetation already prevents > 95% of potential soil loss, consistent with $A_{\max }/A_{\mathrm{updated}}$ around 20.7 (i.e., 1 $-A_{\max }/A_{\mathrm{updated}}$ around 0.95). Consequently, erosion control efforts should focus on the narrow belt of steep sparsely vegetated slopes, where even modest restoration could substantially reduce sediment delivery downstream.

3.2. Revised Soil Loss with NDVI-Based C ($A_{\mathit{updated}}$)

When the NDVI-derived C-factor was substituted for the categorical lookup surface (Figure 2b), 52.5% of all pixels exhibited a positive change in annual soil loss ($\Delta A>0$). For this subset, $\Delta A$ ranged from $8.6\times {10}^{-5}$ to $32.6\mathrm{t}{\mathrm{ha}}^{-1}{\mathrm{yr}}^{-1}$, with a mean of $2.87\mathrm{t}{\mathrm{ha}}^{-1}{\mathrm{yr}}^{-1}$ and a standard deviation of $3.06\mathrm{t}{\mathrm{ha}}^{-1}{\mathrm{yr}}^{-1}$. In contrast, the remaining 47.5% of pixels showed neutral or negative differences, suggesting that the original land cover C values may have underestimated erosion in some degraded or sparsely vegetated zones.

Aggregating over the full model extent yields a baseline sediment export of approximately $95,762\mathrm{t}{\mathrm{yr}}^{-1}$, while the NDVI-refined estimate reaches $151,475\mathrm{t}{\mathrm{yr}}^{-1}$. Therefore, the updated map increases the park-wide estimate by 55,713 t yr⁻¹, corresponding to a relative rise of 58.2%. This upward adjustment is concentrated in transitional vegetation zones and degraded slopes, where NDVI-derived C values are consistently higher than class-based assignments, highlighting the sensitivity of erosion estimates to vegetation detail.

3.3. Maximum Potential Soil Loss ($A_{\max }$)

To assess the worst-case erosion scenario without vegetation or conservation practices, we generated a maximum potential soil loss surface ($A_{\max }$) by setting $C=1$ and $P=1$ across the domain while retaining the calibrated R, K, and $LS$ factors. The resulting map (Figure 2c) highlights steep escarpments and convex hillslopes as the highest-risk zones, with pixel-level losses that exceed the baseline maximum of 285 t ha⁻¹ yr⁻¹, as expected under the bare soil assumption ($C=1$, $P=1$). The area-weighted mean reached 45.4 t ha⁻¹ yr⁻¹, corresponding to an annual sediment export of roughly 3.13 million t yr⁻¹.

Comparing $A_{\max }$ to the NDVI-refined scenario ($A_{\mathrm{updated}}$) reveals that bare soil exposure would produce over twenty-fold greater erosion. Specifically, $A_{\max }/A_{\mathrm{updated}}=20.7$, demonstrating that existing páramo grasslands and their root networks avert more than 95% of potential soil loss. This dramatic reduction underscores the vital role of high-biomass vegetation in maintaining slope stability and regulating sediment flux.

3.4. Erosion Control Index (EC)

The NDVI–based erosion control index (EC, Equation (5)) was mapped for two complete Landsat time series (January–December 2023 and 2024) (Figure 3a,b). Index values ranged from 0 (no protection) to 66.9 (maximum protection under gentle slopes, dense and stable canopy). Following classification (

Table 3. EC classes and areal share in 2023 and 2024. Areas are expressed in hectares; percentages refer to the park’s 28 544 ha. Class thresholds follow FAO (2006) and USDA NRCS (2017).

EC Class	2023		2024
	Area (ha)	Share (%)	Area (ha)	Share (%)
Very low (<0.02)	0.08	0.0003	0.08	0.0003
Low (0.02–20)	28,421.4	99.57	22,597.7	79.18
Moderate (20–40)	120.0	0.42	5939.9	20.80
High (40–60)	0.17	0.0006	79.6	0.28
Very high (>60)	—	—	1.4	0.0049

) five service provision classes were distinguished: very low, low, moderate, high, and very high.

Mean EC across the park increased from 6.3 in 2023 to 13.5 in 2024, while the maximum rose from 44.4 to 66.9 (Figure 3c). High-service pixels cluster on gentle glacier-planed plateaus and in woodlots, whereas very low EC is consistently associated with steep scree slopes, burn scars, and heavily-trodden trekking routes (particularly around the Tres Cruces pass).

The share of the moderate class expanded twenty-fold between the two years (from 0.42% to 20.8%), mainly at the expense of the low class (–20.4 percentage points). This shift mirrors the documented recovery of páramo grasses after the 2023 fire season and a 60% rise in annual mean NDVI (0.05–0.08). Importantly, this observed increase reflects predominantly natural post-fire regeneration; no large-scale planting or earthworks were implemented, although reduced grazing pressure in burned zones likely also facilitated canopy re-establishment.

Combining EC with the updated soil loss map shows that pixels in the highest EC quintile (80 th) experience less than 1 t ha⁻¹ yr⁻¹ of erosion despite being exposed to an $A_{\max }$ of up to 70 t ha⁻¹ yr⁻¹. Conversely, low-EC pixels on slopes > 25% lose 5–15 t ha⁻¹ yr⁻¹. Overall, the vegetation cover present in 2024 is estimated to avert > 95% of the potential soil loss that would occur under bare soil conditions.

The marked improvement from 2023 to 2024 highlights the sensitivity of erosion control to short-term disturbances (fire, grazing) and subsequent vegetation recovery. Areas that remain in the low class are concentrated along tourist corridors and unmanaged grazing hotspots; these constitute prime targets for restoration and visitor-use regulation to safeguard this regulating ecosystem service.

3.5. Hot- and Cold-Spot Clusters

The Getis–Ord $G_i^{*}$ spatial clustering statistic was applied to jointly assess areas of high erosion risk ($A_{\mathrm{updated}}$) and low EC service. This analysis revealed that approximately 5.4% of the park’s area forms statistically significant hot-spots (clusters of high soil loss with insufficient vegetative buffering), while 3.2% of pixels were classified as cold-spots (stable, well-protected zones with low erosion). These clusters are not randomly distributed but rather concentrate along steep escarpments, degraded trails, and overgrazed grasslands, especially near the Tres Cruces pass and the Yanuncay headwaters.

Altogether, the significant hotspots span over 30 km², representing priority targets for erosion control interventions. In contrast, coldspots coincide with glacier-planed plateaus and intact tussock grasslands, offering effective protection even under high rainfall erosivity. These results align with the EC index trends and reinforce the importance of managing slope–vegetation interactions in park planning. Although the park currently lacks a detailed sub-basin delineation for prioritization, the spatial clustering analysis provides a first-order screening of restoration needs based on integrated biophysical and service-based metrics.

3.6. Uncertainty Considerations and Model Robustness

Although this study integrates detailed spatial inputs for erosion modeling and ecosystem-service estimation, a full Monte Carlo uncertainty analysis could not be performed due to the lack of multiple realizations or error surfaces for key input variables (e.g., NDVI, soil parameters, rainfall erosivity). Therefore, we present the main sources of uncertainty in a structured sequence—data sources, preprocessing, and model outputs—in order to clarify their potential influence on our results.

Data Sources: NDVI-derived C and EC surfaces are vulnerable to residual cloud contamination and radiometric inconsistencies across scenes, which can introduce noise in high-mountain imagery and propagate to biased erosion control values. Soil erodibility (K) originates from national-scale polygons rather than field profiles, potentially smoothing local variability in texture and organic carbon. Rainfall erosivity (R) relies on bias-corrected NASA POWER grids downscaled against seven INAMHI gauges; coarse grid resolution may miss sub-daily storm peaks that drive extreme erosion events.

Preprocessing: Each dataset underwent extensive corrections (cloud masking and median compositing for NDVI, kriging for precipitation bias, and DEM void, filling with bilinear interpolation); however, these steps carry their own uncertainties: cloud-masking algorithms can leave artifacts, kriging over complex topography may misrepresent precipitation gradients, and DEM smoothing can underestimate abrupt slope changes. Because RUSLE’s $LS$ factor is nonlinear with slope, even small terrain errors can disproportionately affect modeled soil loss on steep escarpments.

Model Outputs: These input and preprocessing uncertainties multiply through the RUSLE–EC framework. The multiplicative nature of RUSLE amplifies errors in R, K, and $LS$, especially on the steepest slopes. Meanwhile, both RUSLE and the NDVI-derived EC index rest on empirical relationships and spatial proxies that, while robust for identifying relative hotspots, warrant caution when interpreting absolute soil loss values at the pixel scale.

Future work should implement error-propagation or Bayesian approaches, leveraging time series of NDVI, high-frequency in situ rainfall data, and field/sampled soil profiles in order to quantify confidence intervals and conduct rigorous sensitivity analyses across the entire modeling chain.

4. Discussion

This study set out to (i) generate a map of water-driven soil loss for CNP at 30-m resolution using an NDVI-refined rusle workflow and (ii) translate those biophysical outputs into an erosion control (EC) ecosystem service layer that can guide conservation investments. The updated model yields a park-wide mean erosion of approximately $5.3\mathrm{t}{\mathrm{ha}}^{-1}{\mathrm{yr}}^{-1}$, with more than 80% of the predicted sediment originating from only 5% of the landscape, chiefly steep sparsely vegetated escarpments and degraded trekking corridors.

Comparison with an unvegetated scenario, these results show that existing páramo grasslands, shrub mosaics, and scattered Polylepis woodlots prevent about 95% of potential soil loss. A Getis–Ord hot-spot test isolates $30{\mathrm{km}}^2$ where high erosion risk coincides with insufficient vegetative buffering; these clusters represent immediate targets for restoration or stricter visitor regulation.

4.1. Integrating Biophysical Drivers and Ecosystem-Service Delivery

Coupling pixel-scale RUSLE estimates with an NDVI-derived EC index bridges the long-standing gap between hazard mapping and ecosystem service valuation. Unlike categorical look-up tables, the dynamic NDVI surface captures both canopy density (mean NDVI) and phenological stability (coefficient of variation), recognizing that a tussock which browns seasonally offers less protection than year-round evergreen cover. Similar NDVI inversions have cut omission errors in Mediterranean catchments from 25% to below 10% [48], and the 58% upward revision of soil loss estimates reported here confirms that static C factors systematically under-represent degradation on shrub-dominated slopes.

Expressing EC as the proportion of potential soil loss avoided allows managers to weigh biophysical risk against service delivery; plateau grasslands display low absolute erosion yet very high EC scores, whereas scree slopes show the opposite pattern. This dual lens helps to avoid misallocating scarce funds to already stable areas while overlooking small high-risk sites where interventions would yield the greatest marginal gains [49,50].

The EC index we propose here aligns with the regulation service definition of [23], that is, the proportion of potentially lost soil that is actually retained by vegetation. According to [51,52], this biophysical metric represents the supply of the service. Monetizing the associated benefits (e.g., avoided costs of water treatment or dredging) is a subsequent methodological step that remains to be addressed.

4.2. Convergence and Divergence with Recent Literature

The baseline erosion of around $5\mathrm{t}{\mathrm{ha}}^{-1}{\mathrm{yr}}^{-1}$ is an order of magnitude lower than values for unprotected Andean basins dominated by pasture and maize (15–35 t ha⁻¹ yr⁻¹ [53]), underscoring the buffering role of intact páramo. Yet, local maxima exceeding 120 t ha⁻¹ yr⁻¹ rival rates on actively tilled hillsides in the Río Pasto watershed, Colombia [54]. These pockets of extreme loss echo findings from Peru’s Ausangate circuit, where trail widening tripled sediment yield within five years [55].

Replacing categorical C values with NDVI-derived ones increased predicted soil loss by 58%, close to the 40–65% rise reported for semi-arid North Africa after adoption of dynamic C [56]. Year-on-year comparison further shows that a 60% rise in mean NDVI (2023 and 2024) more than doubled the area in the moderate-to-high EC classes; this mirrors post-fire recovery curves in Spain, where EC rebounded to pre-burn levels within two growing seasons [57].

The spatial spillover of erosion costs beyond park limits mirrors broader Andean patterns: landscape-scale modeling in the Colombian Cordillera Central found that loss of native cover inside headwater reserves elevated downstream sediment treatment costs by 30% for municipalities dependent on surface water [54]. Our integration framework similarly highlights the externalities borne by Cuenca’s municipal water utility and the Paute hydroelectric complex, paralleling sediment-hazard concerns around the Río Coca hydropower corridor after the 2020 San Rafael waterfall collapse [58].

4.3. Management and Policy Implications for Southern Ecuadorian Páramo

Hot-spot clusters align with key tributaries supplying Cuenca’s ETAPA water intake system and the Paute hydroelectric complex. Targeted restoration, such as re-vegetation of just the upper 5% of erosion–service mismatch zones or installing elevated boardwalks, could cut modeled sediment export by around 33%, presenting a cost-effective alternative to extensive dredging and water filtration investments [59,60].

With visitor numbers doubling from 2015 to 2023, critical hotspots overlap high-traffic trails around Tres Cruces and Laguna Toreadora. Implementing trail realignment, elevated walkways, or seasonal closures can stabilize exposed soils while preserving recreational access. Similarly, rotational grazing or exclusion of livestock from slopes steeper than 25%, where hoof shear accelerates detachment, will reduce erosion hotspots [61].

Ecuador’s updated Nationally Determined Contribution commits to restoring 500,000 ha of highland ecosystems by 2030 [62,63]. Our NDVI–EC framework provides a scalable and transparent monitoring tool that can guide restoration prioritization in other páramo reserves (e.g., Llanganates, Podocarpus), enhancing the strategic allocation of Green Climate Fund resources.

Results reinforce the emerging view that ecosystem service delivery is place-contingent; identical vegetation greenness may confer vastly different benefits depending on slope and rainfall erosivity. This supports calls to move beyond land cover proxies towards integrative process-based metrics when valuing regulating services [64].

4.4. Methodological Constraints and Sources of Uncertainty

Importantly, the RUSLE–NDVI–EC framework is broadly transferable to other Andean or global mountain systems, provided that three key conditions are met: (1) availability of a high-resolution (≤30 m) DEM for accurate slope–length (LS) calculations; (2) a multi-year time series of cloud-masked surface-reflectance imagery (e.g., Landsat or Sentinel) to derive robust NDVI mosaics and temporal stability metrics; and (3) regionally calibrated soil and rainfall data—either from national soil taxonomy maps and in situ gauges or bias-corrected gridded products—to compute spatially consistent K and R factors. Under these conditions, practitioners can replicate our workflow to identify erosion service hotspots in other steep and data-rich mountainous landscapes, adapting only the NDVI–C regression and EC classification thresholds to local vegetation phenology and management practices.

Three limitations warrant explicit acknowledgment. First, the NDVI-based C factor presumes a monotonic inverse relation between greenness and soil exposure; however, cushion plants and bryophyte mats, which are common above 4000 m, have low NDVI signatures while affording substantial ground protection. Plot-scale calibration of reflectance–cover relations for such physiognomies would refine C estimates [65,66].

Second, soil erodibility (K) parameters were extrapolated from national polygons rather than field profiles, potentially smoothing micro-scale variability associated with histic horizons or volcanic ash layers. Third, the rainfall erosivity surface relies on bias-corrected NASA POWER grids. Although validated against seven INAMHI gauges, kriging may under-represent sub-daily storm bursts that govern mass wasting on escarpments. A Monte-Carlo error propagation approach similar to that applied in Amazonian basins [67,68] would strengthen confidence intervals in future work.

Moreover, as Liu et al. (2020) [49] point out, quantifying the actual benefits requires modeling sediment connectivity to human use points (e.g., water intakes, irrigation), which allows the the biophysical supply (EC) to be translated into on-site and off-site service values.

4.5. Directions for Future Research

Several avenues remain open for extending and strengthening this line of inquiry. First, the temporal dynamics of erosion control should be explored by extending the analysis to the full Landsat–Sentinel archive from 2000 to 2025 [69]. This would allow for the examination of inter-annual trajectories of the EC index, as well as the identification of vegetation recovery or degradation patterns linked to past El Niño disturbances or fire events. Such time series analysis could reveal whether short-term vegetation dynamics significantly alter long-term erosion risk.

Second, sediment fingerprinting using fallout radionuclides (e.g., ¹³⁷Cs or ²¹⁰Pb) is recommended to validate the contribution of modeled erosion hotspots to actual suspended sediment loads [70,71]. This would provide an independent field-based verification of the spatial accuracy of the RUSLE–EC framework and help to distinguish between surface versus subsurface sediment sources in critical watersheds.

Third, future research should incorporate hydro-economic coupling by linking modeled annual sediment yields to the operational costs of dredging reservoirs and treating turbidity at water intake facilities [72]. Such integration would enable a monetary valuation of erosion mitigation, demonstrating the direct financial benefits of maintaining high-EC zones for downstream water security.

Fourth, climate change scenarios should be explicitly incorporated by downscaling CMIP6-based rainfall erosivity projections and modeling prospective land use changes such as road expansion or grazing intensification. These scenarios would help to determine whether natural vegetation recovery alone can compensate for increased climatic pressures or whether more active restoration will be required [73,74].

Lastly, participatory co-design approaches should be prioritized in order to engage local stakeholders, particularly herders and park guides, in developing feasible interventions. Collaboratively identifying alternative grazing zones, trail realignments, or re-vegetation protocols can increase the social acceptability and long-term success of conservation efforts, ensuring that scientific recommendations align with on-ground realities.

In sum, integrating an NDVI-refined rusle model with an erosion-control ecosystem-service lens reveals that while CNP largely retains its sediment-buffer function, small spatially-clustered hotspots jeopardize downstream water and energy security. The framework presented here offers a replicable blueprint for aligning conservation practice with national restoration pledges and climate resilience goals across the tropical Andes.

5. Conclusions

Intact Andean vegetation remains the cornerstone of slope stability in El Cajas. Although substituting static land cover C factors with a continuous NDVI-derived surface raised the park-wide soil loss estimate by 58%, existing páramo grasslands, shrub mosaics, and Polylepis woodlots still avert over 95% of the potential erosion that would occur under bare soil conditions. This underscores that preserving and restoring high-biomass cover is by far the most effective means of safeguarding soil and downstream water–energy services.

By coupling RUSLE outputs with an NDVI-based erosion control index (EC) that accounts for canopy density, seasonal stability, and slope, we demonstrate both the vulnerability and resilience of these ecosystems. A 60% increase in mean NDVI between 2023 and 2024 more than doubled the area offering moderate-to-high protection, revealing rapid natural recovery after disturbance. Yet, just 5% of the landscape (approximately 30 km² of steep sparsely vegetated terrain) produces 80% of sediment export. Directing restoration or visitor management measures to these hotspots could cut overall sediment loads by a third, delivering outsized ecological and economic returns.

As a result, ecosystem-based interventions in these hotspots may be more cost-effective than downstream dredging and filtration; however, quantifying monetary savings requires site-specific unit cost data, and is beyond the scope of this study.

The integrated RUSLE–NDVI–EC framework offers a transferable remotely-sensed tool to quantify the biophysical foundation of erosion control ecosystem services, specifically, the avoided soil loss due to vegetation cover. Although this study does not explicitly estimate the economic or social value of this regulation, it provides spatially explicit indicators that can inform future valuation efforts and guide conservation investment prioritization. It aligns directly with Ecuador’s national restoration targets, and can inform hydro-economic analyses that monetize the benefits of maintaining high-EC zones under future climate and land-use scenarios.

While this study robustly quantifies the supply of the soil erosion control service (EC), its economic valuation and estimation of societal benefits remain a priority for future applied research. The proposed workflow is directly applicable to other mountain regions where a fine-scale DEM, periodic satellite imagery, and locally calibrated soil/rainfall data are available, enabling rapid data-driven prioritization of erosion control services globally.