The peer-reviewed literature search was conducted using the Scopus bibliographic database, and targeting studies published between 1980 and 2020. We searched within the article title, abstract and keywords for the following terms: ∗erosionORflood∗ORlandslid∗ORmassmovementORalluv∗ORrunoffORinfiltrationORgullyORsedimentORdepositionORsoil AND AndesORColombiaORVenezuelaOREcuadorORPeruORBoliviaORChileORArgentina AND ∗forest∗ORgrazingORgrass∗ORpastureORagricultureORcrop∗ORlanduseORpunaORparamoORbofedal∗ORdamORreservoirORconservationORmanagementORtill∗ORterracesORirrigationORlake∗ORhydraulic∗ORancientknowledgeORarchaeologyORhumanORpeopleORanthropogenic For the gray literature, we searched in 35 different databases from specialist organizations, public institutions and local repositories of private and public universities in the Andean region. The abovementioned search criteria were adapted for the gray literature given the limited search capabilities of some of the databases. Full details on the literature search are provided in the Supplement B, including the complete search terms, the number of records generated for specific searches and the name, location and search dates in Scopus and the national and regional databases of research institutions, universities and specialist organizations. For international peer-reviewed literature, we used a test library of 20 references (Supplement E) that confirmed that the search strings captured relevant literature.

The number of studies that were identified, screened, selected and included in the analysis are shown in a PRISMA flow diagram (Fig. 1). Between 10 January and 27 February 2020 we identified 1798 potentially relevant studies: 91 % corresponding to peer-reviewed articles and 9 % to gray literature. After removing duplicate studies, the dataset was reduced to 813 studies. These records were screened and articles that fulfilled the following criteria were included in the database: (1) they present quantitative data on soil erosion or soil quality comparing sites with different land use and protective land cover, soil and water conservation measures or elements of hydraulic regulation, (2) they are experimental studies including observational datasets or are modeling studies that are fully validated with field experiments or measurements and (3) they were realized in the Andean region. During the screening stage, we excluded 623 studies because of absence of quantitative on-site or off-site soil erosion or soil quality measurements.

We assessed 190 studies in full-text and further excluded 54 papers as the studies did not report quantitative measures of erosion rates or soil quality for different classes of land use and protective land cover, soil and water conservation measures or elements of hydraulic regulation. At this stage, this mainly concerned scientific reports on landslides and landslide-related erosion events.

A total of 136 studies were included in the systematic review. Where a study encompassed several independent case studies, the case studies were included in the final database as separate entries. Each case study was coded by a unique study identifier and recorded in the georeferenced database (Supplement C). We recorded the following ancillary geographic data: (1) country, (2) site name, (3) coordinates (latitude and longitude in decimal degrees), (4) elevation (meters above sea level, m a.s.l.), and information on (5) bioclimate, (6) surface lithology, (7) ecosystem and (8) landform. The latter four variables were derived from the 2005 Nature Conservancy datasets via the USGS dataviewer for South America (https://rmgsc.cr.usgs.gov/, last access: 27 April 2021). We included additional information on the type of study: (8) the experimental design following the classification scheme of Nichols et al. (2011), (9) the modeling approach based on a classification in statistical, process-based and mixed models, (10) the existence of field data and (11) the spatial scale and organization of the study based on a classification in plot (<0.01 km2), small catchment (between 0.01 and 1000 km2), large catchment (>1000 km2), and landscape scale (regional) analyses. The latter contained data collections that are not organized by hydrological units and that include measurements taken over a larger geographical area.

In the analyses, we quantified the effect of restoration and protection of natural vegetation, such as forest or native grasslands (PRO), forestation with native or exotic species (FOR) and implementation of soil and water conservation measures (SWC) for soil erosion and mitigation (Fig. 2). Soil and water conservation measures (SWC) include crop management, conservation tillage and slow-forming terraces and the implementation of linear elements, such as vegetation strips and check dams. We compared the three natural infrastructure interventions (PRO, FOR and SWC) with untreated areas under traditional agriculture, either cropland (CROP) or rangeland (RANGE), and bare land (BARE). Bare land corresponds to abandoned cropland or degraded land with very low (<10 %) vegetation cover.

The erosion indicators included in this study were (Fig. 2): soil loss rate (Sloss), determined as soil loss in t km-2 yr-1; plot run-off coefficient (RC), determined as event-based run-off coefficient from rainfall simulation experiments, in %; specific sediment yield (SSY), determined as the catchment-wide sediment yield per surface area in t km-2 yr-1, in abandoned cropland or degraded land with very low (<10 %) vegetation cover; and catchment-wide run-off ratio (RCC), determined as the annual total run-off ratio of the catchment, in %. While Sloss and SSY are direct measures of soil erosion at the plot and catchment scale, the plot and catchment-wide run-off coefficients (RC and RCC) are indirect indicators of soil erosion by water: the rainfall regime plays a role as raindrop impact and run-off water are involved in the detachment of soil particles and transport of sediment in surface water flow. Empirical studies compiled by, for example Bonnesoeur et al., 2019 and Valentin et al., 2008 have shown the strong association between run-off coefficients and soil erosion rates.

In addition to the four erosion indicators, two soil quality indicators were included: SOC (total soil organic carbon of the uppermost soil horizon, between 5 and 30 cm, in %), and BD (dry bulk density of the topsoil horizon, between 5 and 30 cm, in g cm-3). The SOC is the main indicator of soil quality (Franzluebbers, 2002) and is directly linked to key soil functions (Wiesmeier et al., 2019) including soil water retention, erosion prevention and resilience to drought and floods (Paustian et al., 2016). The BD is a commonly reported soil physical property that is related to soil aeration, water and air permeability and soil microporosity (Horn et al., 1995). Increased bulk density can be indicative of soil compaction and affect the water retention capacity and accelerate soil erosion (Molina et al., 2007; Patiño et al., 2021). Other erosion indicators were recorded in the database but not included in the statistical analyses because of a lack of statistical representation. These include plot-based indicators like the stock in SOC over the entire soil depth or the saturated hydraulic conductivity of the topsoil or catchment-wide indicators like the presence or relative occurrence of erosion signs or the suspended sediment concentration in the river channels. Mean, sample size and deviation metrics were extracted from figures using PlotDigitizer. Information from in-text tables and supplementary material was copied and tabulated in spreadsheets.

Of the 136 studies included in the systematic review, 118 studies contained sufficient information on the soil erosion and soil quality indicators to be statistically analyzed. Besides the abovementioned information, the georeferenced database includes bibliographic details and a URL link to the individual case studies (Supplement D).

First, we tested whether sites with natural infrastructure interventions (PRO, FOR and SWC) are different in on-site (Sloss, RC) and off-site (SSY, RCC) soil erosion and soil quality (SOC, BD) compared to untreated areas under traditional agriculture (CROP, RANGE) or bare land (BARE) as illustrated in Fig. 2. The comparison of the four erosion and two soil quality indicators between the treatments was performed using one-way analysis of variance (ANOVA). In this analysis, we pooled all observations from the 118 case-studies. Because of the limited number of quantitative case studies for the Andes, the number of observations is not the same for each group. Given the low number of observations per group, the Kruskal–Wallis ANOVA on the ranks was applied, with Dunn's post hoc test. We rejected the null hypotheses (i.e., that there are no differences between the means of the groups) at the 0.05 significance level. We used R (R Core Team, 2021) with the “PMCMRplus” package (Pohlert, 2018) in R to perform the non-parametric comparisons.

Next, we analyzed the overall effect of natural infrastructure interventions on soil erosion and soil quality indicators. In this analysis, we only included case studies with a control–treatment design, where quantitative measures of soil erosion and quality were available to establish the control–treatment contrast. The response ratio (RR) was then used to determine the effect sizes. In this study, the RR was calculated for each natural infrastructure intervention (PRO, FOR, SWC) and soil erosion and quality indicator (Sloss, RC, SSY, RCC, SOC, BD). For the control group, we combined data of sites with traditional agriculture, either cropland (CROP) or rangeland (RANGE) and bare land (BARE) given the limited number of matched pairs of control and single or multiple treatment(s). For each pairwise comparison, we plotted the effect size of the individual studies in forest plots and explored the heterogeneity in the response among the case studies. These plots were used to identify the magnitude and sources of variation among the studies and to identify possible outliers. We then extracted the central tendency (mean effect) and confidence limits (standard error) for each indicator and pairwise comparison. The mean effect and its standard error were plotted in summary forest plots (per pairwise comparison) to assess the overall effectiveness of a specific intervention on soil erosion and quality indicators. The graphs were produced using the R-package “metafor” (Viechtbauer, 2010).

Of the 118 studies evaluating the effect of natural infrastructure interventions on soil erosion and quality indicators, 54 studies contained data on soil and water conservation practices (SWC), 50 studies on protective vegetation (PRO, FOR or both) and 14 studies on all 3 (SWC, PRO and FOR). The majority of studies were journal articles (79 %), followed by gray literature (14 %) and chapters from books (7 %). The studies covered a 6500 km long stretch across the Andes, with 4 % of the studies in Venezuela (n=5), 6 % in Colombia (n=7), 36 % in Ecuador (n=43), 35 % in Peru (n=41), 7 % in Bolivia (n=8), 8 % in Chile (n=9) and 4 % in Argentina (n=5). Ecuador and Peru had the highest number of case studies (Fig. 3). The large majority, i.e., 89 %, of the studies investigated soil erosion in tropical climates, with 59 % of the studies performed in pluvial seasonal, 19 % in pluvial and 10 % in desertic or xeric climates. The remaining studies were performed in temperate or Mediterranean climate regimes. Field studies mostly involved erosion measurements at the plot scale (48 %, n=57), small catchment scale (18 %, n=21), and landscape scale (22 %, n=26). Only 12 % of the studies included erosion assessment at the scale of large catchments (>1000 km2).

The one-way analysis of variance revealed significant differences between treatment and control in soil quality and on-site soil erosion with notable differences in SOC (p<0.01,n=85), BD (p=0.02, n=46), soil loss (p<0.01, n=123) and plot run-off coefficient (p=0.03, n=37) (Table 1; Fig. 4). Notably, none of the erosion indicators that were measured at the catchment scale were significant at the 0.05 level, as we observed only small differences between categories for specific sediment yields (p=0.10, n=37) and no differences for catchment-wide run-off coefficient (p=0.59, n=18). The latter might be due to the limited number of observations documenting the effect of natural infrastructure interventions on SSY (n=37) or RCC (n=18) and inherent variability in run-off and sediment discharge at the catchment scale as shown by Tolorza et al. (2014) and Molina et al. (2015). Below, we only present tendencies that are statistically significant at the 0.05 level.

The SOC concentration of topsoil is the indicator with the highest significance level, showing strong differences in soil quality between protected sites, cropland and bare soil (Table 1, Fig. 4). Based on 85 observations, we observed soil organic carbon concentrations of the topsoil between 0.47 % and 34.06 %, with mean values of 4.47 % ± 0.62 %. Based on the results of the post hoc Kruskal–Wallis rank sum test, two distinct groups can be identified: (1) areas covered by natural vegetation such as forests and native grasslands with a mean SOC value of 8.67 % ± 1.89 %, and (2) areas covered by agricultural crops and bare land having a mean SOC value of 2.49 % ± 0.32 % and 1.88 % ± 0.49 %,respectively. Areas with SWC measures belong to the second group with a SOC of 2.96 % ± 0.70 %. Rangelands and plantation forests have intermediate SOC values of 6.21 % ± 2.05 % and 3.17 % ± 0.71 %, respectively. These results are consistent with the systematic review of Bonnesoeur et al. (2019) that reported lower levels of topsoil organic matter in plantations compared to native forests and grasses. However, the differences reported here are not statistically significant (p=0.12).

Soil BD of the topsoil is reported in 15 % of the reported case studies for different natural infrastructure interventions. Soil BD ranges between 0.36 and 1.67 g cm-3 with a mean value of 1.07±0.05 (Fig. 4; Table 1). The lowest mean BD values, i.e., 0.82±0.08 g cm-3, are observed in soils covered by native vegetation. Although the mean BD values are notably higher in areas with cropland, forestation, rangeland, and particularly bare land, the wide range in reported BD values per category does not allow us to distinguish them from areas covered by natural vegetation at the 0.05 significance level. Remarkably, areas under SWC treatment have significantly higher BD values compared to natural vegetation (Table 1), which might reflect the advanced state of physical soil degradation due to compaction before SWC intervention (e.g., Rymshaw et al., 1997). It also highlights that it may take several years to decades for impacts to be reversed and that high levels of subsurface compaction may be irreversible without soil restoration (Borja, 2018).

The rate of soil loss measured at the plot scale (t km-2 yr-1) is one of the most common indicators of soil erosion, as it is reported in 43 % of the case studies. The 125 quantitative measurements of Sloss reveal that Sloss rates vary widely with mean value of 2590±295 t km-2 yr-1 and minimum and maximum values of 0.001 and 14761 t km-2 yr-1, respectively. Significant differences in Sloss are observed between areas covered by natural vegetation and crop or bare land, with soil losses being on average 11 to 18 times lower in areas with protected vegetation (Table 1). Rangelands, areas with forestation and SWC measures have intermediate values of Sloss (2370, 1860 and 1660 t km-2 yr-1),respectively, and are not significantly different from natural vegetation, crop or bare land.

The run-off coefficient (RC) is measured as surface run-off at the plot scale, and is here reported as the percentage of the rainfall that becomes run-off. The number of case studies that report run-off coefficients for different categories of natural infrastructure is low (12 %). Figure 4 illustrates the wide range of RC values (min: 0 %, max: 47 %) that are observed in the Andes, with mean values of 12.9±2.23. The large variation might be the result of inherent spatial heterogeneity in rainfall-run-off response (Guzman et al., 2019). However, methodological bias cannot be excluded as multiple field methods to estimate plot RC were used: portable rainfall simulators covering a few cm2 (e.g., Harden, 2001), runoff plots covering 1 m2 (e.g., Perrin et al., 2001; Molina et al., 2007), and experimental sites covering >10 m2 (Molina et al. 2009; Suescún et al. 2017). Also, the amount and intensity of the (simulated) rainfall often vary between case studies. Notwithstanding, significant differences are observed in RC between areas with soil and water conservation measures and bare land, with RC values being on average 3.5 times lower in SWC compared to BARE (Fig. 4).

When limiting the quantitative analysis to matched pairs of control and single or multiple treatment, the number of independent empirical studies is reduced from 118 to 89. For the analysis of the response ratios, the sites with traditional agriculture, either cropland (CROP) or rangeland (RANGE), and bare land (BARE) were regrouped into one control group. This is justified by the fact that the soil quality and erosion indicators are not significantly different between the three types of control sites (Table 1). Figure 5 and Table 2 show the effect size of (i) restoration and protection of natural vegetation, (ii) forestation, and (iii) SWC on SOC, BD, RC, Sloss rate, SSY and RCC. Below, we only discuss results that are based on a minimum of 4 independent treatment-control studies.

Amongst the three intervention types, the protection and conservation of natural vegetation (PRO) has the strongest effect on soil quality (SOC, BD) and erosion (Sloss, SSY). When native forests and grasses are protected from conversion to agricultural land, the topsoil contains 3.01±0.893 times more SOC than in the control sites. At the same time, the soil physical structure is better with a dry BD of 0.82±0.08 g cm-3 in natural vegetation, being 0.878±0.030 times lower compared to control sites. The high soil porosity enhances structural support, water and solute movement and soil aeration (Podwojewski et al., 2002). There is a clear and positive effect on soil erosion mitigation, with Sloss being 0.357±0.187 times lower, and SSY being 0.334±0.085 times lower than at the control sites. Experimental work by, for example, Janeau et al. (2015) showed the importance of native vegetation in facilitating soil water infiltration as it can conduct over 50 % of rainwater through stemflow to the soil. This is confirmed by other empirical data (e.g., Harden, 1996; Poulenard et al., 2001) collected at the plot scale, and the RC is on average 0.318±0.159 times lower in areas where natural vegetation is protected or conserved. The empirical data are not sufficient to systematically assess the effect on run-off processes at the catchment scale.

Only 17 % of the records on treatment-control experiments contain information on the effect of forestation with native and/or exotic species (FOR) on soil quality, on-site and off-site soil erosion (Table 2). The database counts less than 3 empirical studies on rainfall-run-off generation, at the plot and catchment scale. Compared to the control sites, a positive effect is reported on soil quality, with 1.19±0.125 times higher SOC and 0.959±0.032 lower BD. The pairwise analysis did not show evidence of a net effect of forestation on soil erosion (Sloss): the response ratio shows large scatter with RR values ranging between 0.37 in the study by Henry et al. (2013) and 7.75 in the case published by Pesantez and Seminario (2010). Notwithstanding the high variability in response ratios for on-site erosion (Sloss), a positive effect was observed for the catchment-wide sediment yields with SSY being on average 71.3 % ± 27.7 % of the yields measured in control sites. Similar observations were made by Bonnesoeur et al. (2019) who attributed the scatter in the empirical studies to the type of forestation (native vs. exotic species) and forestation age. In addition to this, the prior state of the environment (soil quality and erosion) has a major impact on erosion mitigation as Balthazar et al. (2015) showed for a case in the Ecuadorian Andes.

Almost 50 % of the treatment-control studies concern interventions with soil and water conservation measures (SWC). There is a net positive effect of the implementation of conservation measures on the soil organic carbon content of the topsoil, with values that are 1.28±0.170 times higher compared to control sites (Fig. 5). The effect on the BD is small, with BD in treated sites being 94.6 % ± 3.2 % of the values measured in control sites. The limited effect on soil BD suggests that the recovery of the soils' physical structure from compaction is slow, even within the topsoil (Jacobi et al., 2015). Soil loss rates changed significantly after the application of SWC measures: Slosses are reduced to 62.1 % ± 9.2 % of their original values after the intervention. The effect of infiltration ditches on soil erosion mitigation is particularly well-documented for the Peruvian Andes, where Vasquez and Tapia (2011) reported soil erosion rates that were more than two times reduced after the intervention (i.e., from 4500 to 2060 t km-2 yr-1). The effect is strongest when the measures are applied on abandoned cropland or degraded land with very low (<10 %) vegetation cover (Fig. 3), as shown by De Noni et al. (2001) in various case studies distributed along the Ecuadorian Andes. In contrast to the response ratios for Sloss, the plot-scale RC shows large scatter with values of 0.740±0.316. The scatter can be attributed to strong differences in hydrological response between control sites, with degraded and abandoned land generating more run-off than rangeland or arable land (Molina et al., 2007).