We created a detailed inventory of bedrock, block- and talus slopes (area ≥ 0.001 km²) within five key sites of the ANC using satellite imagery (Esri, Google Earth Pro) and a TanDEM-X DEM with a 12 m resolution (DLR, 2017) (see Table 1, Supplementary Table S3). Satellite images from austral summer months with minimal snow cover and shading were selected to enhance mapping accuracy. All five key sites are located within the periglacial belt (regional lower permafrost limit ≥3,700 m asl, Trombotto, 2000; PZI, Gruber, 2012) (see Figure 1B, see Supplementary Figure S4). The lower limit of 3,700 m is a conservative estimate to encompass the entire regional periglacial belt, though other sources specify limits between 3,900 and 4,000 m asl (Brenning, 2005; Schrott, 1996; Croce and Milana, 2002). The selected key sites span different altitudes, slope angles, valley orientations, and proportions of inventoried cryospheric landform cover to adequately represent the spatial heterogeneity of the ANC (see Table 1; Figure 1). In total, they account for 160 km², corresponding to ∼24% of the study area potentially underlain by permafrost (PZI, Gruber, 2012) and 12.2% of the entire ANC. The key sites serve as (1) example regions to illustrate regional topoclimatic, geomorphologic and cryospheric conditions, and associated processes, (2) manual mapping areas, and (3) calibration and validation areas for the three different modeling techniques.

Mapping results from the first two manually mapped key sites (ANC, SLC) were revised during a field trip in February 2022. Field observations were consistently applied in the subsequent mapping (Olivares, Cerro Pata, Cordón de las Minitas). (Debris-covered) glaciers, rock glaciers and perennial snowfields (area ≥ 0.01 km²) from the Argentinean national inventory of cryospheric landforms (IANIGLA-CONICET, 2018), as well as manually mapped bedrock were masked. The national inventory’s restricted rock glacier delineation excluding frontal and lateral margins (RGIK, 2022), was maintained to ensure model transferability at regional scales. However, we consider the margins of rock glaciers as part of the landform and intentionally left these areas unmapped. A visualization of the mapping results in the three key sites ANC, SLC, and Olivares can be found in Köhler et al. (2024). We mapped only clearly distinct landforms and set the size threshold for consistent and transferable mapping results. Thus, the inventoried landform proportions slightly underrepresent their actual distribution in the ANC. However, the high level of detail of the geomorphological mapping, its validation by three co-authors, and the use of the national inventory (IANIGLA-CONICET, 2018) strongly reduce the uncertainties in the input dataset and thus in the predictive modeling.

We used 15 predictor variables derived from the TanDEM-X DEM (void-filled, resolution: 12 m, DLR, 2017) for predictive modeling. The predictors directly or indirectly represented topography, surface morphology, climatic conditions, and contributing area characteristics across different spatial scales (see Table 2, Supplementary Table S3). They characterize the spatial distribution, topoclimatic conditions and geomorphic properties of block- and talus slopes in the ANC (Köhler et al., 2024). For more information, please refer to the companion paper.

By using sine and cosine transformations, circular aspect variables were decomposed into the directional components “north-exposedness” and “east-exposedness” (Brenning and Trombotto, 2006; Brenning, 2009). Climate and moisture conditions are represented by DEM-derivatives due to a lack of measured data in the catchment. Potential incoming solar radiation (PISR) was calculated as the annual sum for 2022. The size of the contributing area (D8-flow algorithm) was log-transformed due to its skewed distribution and wide range of values (Brenning, 2009). The scaling of environmental predictors entering the predictive models is highly dependent on the landform morphometry and its controlling environmental conditions. Thus, we applied an automated scaling procedure (Sîrbu et al., 2019) to identify optimal scaling for slope inclination (slope), topographic position (TPI), and curvature (overall, planform, profile) (see Table 2, Supplementary Table S3). All analyses were conducted in Esri ArcMap (10.4) and open-source GIS and data analysis tools, including QGIS (3.10.9), SAGA GIS (9.3.0), and RStudio (4.3.2).

The quality of the input dataset has a high impact on model quality, robustness and transferability. Various prerequisites were addressed during data preparation: (1) The input data needed to reflect the heterogeneity of the study area (Schoch et al., 2018). (2) the mapped inventory of block- and talus slopes was converted into binary rasters, where “1” indicated landform presence and “0” denoted absence. Each pixel was assigned with the site-specific characteristics of the 15 terrain attributes. (3) (Multi-) collinearity between predictor variables reduces model accuracy and complicates interpretation, as the individual effect of correlated variables on the response variable cannot be determined accurately (James et al., 2013; Heckmann et al., 2014). Thus, geostatistical upscaling requires non-collinearity among predictors, and we computed the variance of inflation factor (VIF) to identify (multi-)collinearity within our predictor set (Aguilera et al., 2006; James et al., 2013). (4) We used different models to predict the probability of a pixel being classified as 1 (blockslope/talus slope presence) or 0 (blockslope/talus slope absence) based on stratified random samples from all key sites. Due to the high blockslope occurrence in the ANC, a ratio of 1:1 (landform presence vs absence) was implemented for each random sample (Brenning, 2005). The aerial share of talus slopes is lower, and thus an even ratio of landform presence and absence in the samples might overestimate the talus slope distribution at the expense of other landforms. Heckmann et al. (2014) note that biases toward small probabilities may occur if the ratio of events to non-events strongly overrepresents one of the cases in the samples. To avoid this and to better reflect the actual conditions, we created random samples with a ratio of 1:3 for talus slopes.

We compared three different classification algorithms to analyze the spatial distribution of target landforms: logistic regression (LR), generalized additive models (GAM), and random forest (RF).

Sample size can strongly influence model robustness (i.e., reproducibility) and comparability due to over- or underfitting, spatial autocorrelation, and overparameterization (Heckmann et al., 2014). To determine the ideal sample size, we conducted a pixel-based sample size analysis for all three models (number of pixels per sample = 100, 500, 1,000, 2,000, 4,000, 8,000, 16,000, 32,000, 48,000), maintaining predefined ratios of landform presence and absence for both target landforms (1:1 for blockslopes and 1:3 for talus slopes). The optimal sample size was identified based on normalized changes in the area under the receiver operating characteristic curve (AUROC) values and their interquartile ranges (IQR) across ten model runs with 15-fold validation. The ideal sample size was specified by the class above which model improvements became insignificant (defined here as <0.06) or model performance impaired.

We evaluated the predictive power, robustness, and spatial transferability using both spatial and non-spatial validation (Brenning, 2009). Each classification algorithm was trained and tested in 10 model runs with 15-fold validation each, using the ideal sample size determined from the sample size analysis and the predefined landform presence and absence ratios. Unique stratified random samples from all key sites were selected without replacement. We evaluated and compared model performances based on AUROC values, IQR, and model accuracy (sum of the true positive and true negative rate). We transferred probabilities of landform occurrence into binary classes of presence (1) or absence (0) using the optimal cutoff derived from AUROC values, which was further examined for geomorphic plausibility (Heckmann et al., 2014; Steger et al., 2016). The AUROC scores the model performance based on sensitivity (true-positive rate) and specificity (true-negative rate), and ranges from 0.5–1.0 (Brenning and Azócar, 2010). Higher AUROC values signify higher predictive power, with values >0.9 implying an excellent, 0.9 to >0.8 a good, 0.8 to >0.7 a fair, and ≤0.7 a poor model (Araujo et al., 2005). Common non-spatial validation approaches based on splitting one dataset into training and testing subsets have limitations due to dependencies between datasets (Brenning, 2009; Schoch et al., 2018). Therefore, we additionally performed spatial validation (10 model runs, 10-fold validation each) and assessed the geomorphic performance and plausibility in independent test areas outside the training domain (Steger et al., 2016; 2021).

The model whose predictions most closely matched the actual landform distribution (predictive power) and provided consistent results across all model runs both inside (robustness) and outside (spatial transferability) the training domain was identified as the best performing model. This model was then applied to the area above the regional lower permafrost limit (≥3,700 m asl) in the ANC (model domain = 686 km²).

Furthermore, we analyzed the environmental conditions and formative controls of block- and talus slopes based on predictor importance using variable selection frequency (LR & GAM) and relative predictor importance (RF). We evaluated the effect of the influential variables on predicted probabilities to identify key characteristics of block- and talus slopes in the study area.

All models demonstrate good to excellent predictive performances, with mean AUROC values ranging from 0.83 to 0.96 (blockslopes) and 0.85 to 0.95 (talus slopes), along with mean accuracies of approximately 0.82 (blockslopes) and 0.85 (talus slopes) based on non-spatial validation (Araujo et al., 2005) (see Figure 3, Supplementary Table S2). This confirms the model’s ability to distinguish between the presence and absence of the target landforms. Overall, the predictive performance of talus slope modeling is slightly lower and characterized by higher uncertainty compared to blockslope modeling when applying RF and GAM, while the opposite is true for LR.

For both landforms, the predictive ability and geomorphic plausibility of RF clearly outperforms LR and GAM, with LR being least suited in model comparison. The Kruskal–Wallis rank sum test demonstrates statistically significant differences in model performances (p < 0.05) (Schoch et al., 2018). Thus, RF is the most suitable classification algorithm for our predictive modeling approach with small uncertainties, high robustness and good model performance (high AUROC, low IQR).

Increasing sample sizes lead to improvements in AUROC values and reductions in IQR (see Figure 4, for full sample size analysis see Supplementary Figure S1). With the lowest mean AUROC values starting at ∼0.86 and ∼0.82 for block- and talus slopes, respectively, RF performs well even with small sample sizes. However, the smallest sample size class (n = 100) exhibits a comparatively wide IQR (>0.3), indicating lower model robustness and transferability. As sample size increases, AUROC values follow a near-linear progression but show a slight dip at the largest sample size class. For smaller sample sizes (n ≤ 2,000), talus slope modeling yields higher uncertainties and lower predictive power compared to blockslope modeling. However, performance measures converge at sample sizes >2,000. Maximum mean AUROC values of ∼0.96 for both landforms and minimum IQR (≤0.0015) are reached at n = 48,000. The ideal sample size was concordantly established to be 32,000, beyond which model performances hardly improve or possibly even decrease due to overfitting (mean AUROC >0.95, IQR ≤0.0017).

Modeled block- and talus slope coverage revealed that both landforms account for almost 79% of the area above the regional lower permafrost limit (3,700 m asl) and exhibit a distribution ratio of ∼61% (blockslopes) to ∼17% (talus slopes), similar to the key sites (see Figure 8A; Table 3). Overlapping predictions affect approximately 5% of the modeled block- and talus slope area, primarily in the eastern, lower part of the ANC towards the lower permafrost limit and thus the model boundary (see Figure 8B). Elsewhere, the overlaps are minimal and mainly confined to transitional zones between block- and talus slopes in middle slope positions and areas with small to medium-sized bedrock outcrops. The remaining 21% of the model domain, not covered by either the target landforms or the inventoried cryospheric landforms (IANIGLA-CONICET, 2018), is occupied by other landforms, including bedrock, alluvial fans, and floodplains, which are beyond the scope of this study. These features are correctly excluded from both models, while frontal and lateral aprons of rock glaciers, which are not included in the national inventory (restricted rock glacier delineation, see Section 3.1) are typically interpreted as talus slopes. The model slightly underestimates the distribution of blockslopes compared to manual mapping, whereas no clear: trend is observed for talus slopes (see Table 3). Only in Cordón de las Minitas, both models overestimate the respective landform distributions, while the largest deviation occurs at the Olivares key site. On average, the predictive models exhibit a negative offset of −3% for blockslopes and a slightly positive offset of 0.5% for talus slopes. However, the overlap has to be considered when interpreting these values.

The relationship between the most influential predictor variables in the RF model (see Section 4.1.2) and the predicted probability of block- and talus slope occurrence highlights heterogeneous distribution conditions for both landforms. General trends rather than specific statements can be drawn from this. Most predictors show non-linear relationships with both landforms (see Figure 9, for all variables see Supplementary Figure S3).

Blockslopes occur at all modeled elevations and at various topographic positions but become more probable with higher elevations and slope positions. Peak probabilities around 0.8 at neutral to moderately positive TPI₁₀₃ values confirm this and indicate their primary occurrence on middle to upper, open slopes with constant inclination or low-relief areas. Blockslopes show a bimodal distribution at lower slope angles (≤15°), displaying either very high or very low predicted probabilities. The maximum likelihood of blockslope presence is observed at 30°–35°. Above this value, the probability declines, making their classification highly unlikely at slopes >45°. Blockslopes have minimal dispersion around zero for the overall curvature (curv₁₃₉ ± 0.0015), suggesting they are vertically and horizontally elongated with minor lateral convexity and upward concavity. Their highest predicted probabilities are associated with low roughness values (TRI) < 10. Potential incoming solar radiation (PISR) has a considerable effect on the prediction of blockslopes. Blockslopes are rare and highly unlikely below a PISR of ∼1.3 MWh/m² but become significantly more frequent above this value. However, occurrence probabilities exhibit high dispersion. Blockslope probability only begins to decline at approximately 2.4 MWh/m². In contrast to talus slopes, however, they continue to occur frequently and exclusively with high likelihoods.

Talus slopes are predominantly located at foot slopes towards the valley bottom or at cliff bases, with the highest predicted probabilities at low TPI₁₁ values. TPI₁₁ ∼0 indicate their occurrence on open slopes with constant inclination or flat areas (Weiss, 2001). These locations are primarily in proximity to major bedrock outcrops, below large headwalls and within the rooting zones of inventoried rock glaciers (IANIGLA-CONICET, 2018) (see Figures 8, 9). Supporting this, talus slopes reach highest probabilities between 3,700 m asl (lower limit of predictive modeling approach) and ∼4,000 m asl. With increasing altitude, they become less likely, rarely existing above 5,000 m. Talus slopes display low sensitivity to slope angle, primarily occurring between 20°–40° in the study area. Overall curvature ∼0 confirms that talus slopes are vertically and horizontally planar, thus slope angles vary only slightly. They are mostly distributed at TRI values <10, with a few occurrences at higher values. There, they are more likely to occur than blockslopes. In general, talus slopes are less sensitive to the amount of incoming solar radiation and occur across all recorded PISR quantities. Yet, the parameter is ranked with high importance by the RF model, and a negative correlation is observed beyond a potential annual budget of ∼1.8 MWh/m². With increasing PISR, the likelihood of talus slope formation declines to almost 0% above ∼2.3 MWh/m².

The consistently high predictive performances of the RF model with low uncertainties suggest that RF effectively captures the distributional characteristics of block- and talus slopes in the ANC. The significantly poorer performance of LR compared to GAM and RF (see Figure 3) demonstrates that the widespread, heterogeneous distribution of talus slopes, and even more so of blockslopes, requires modeling techniques that can cope with these heterogeneous, often non-linear relationships. The superior model performance of RF in the sample size analysis (see Supplementary Figure S1) highlights the advantages of data-driven over more model-driven classification algorithms when working with complex, non-linear data, although a thorough examination of model plausibility is required (Breiman, 2001; Schoch et al., 2018; Steger et al., 2021). However, achieving good to excellent model performance requires careful consideration of various prerequisites and limitations:

1. Addressing mapping uncertainties: Geomorphological mapping based on remotely sensed imagery is subject to uncertainties. These should be minimized to ensure high accuracy in predictive modeling. Some studies demonstrate large mapping heterogeneities of up to 30% due to variable image resolution and quality, inter-operator mapping style, and a lack of established guidelines for landform classification and mapping, with the latter two appearing to produce the highest uncertainties (Brardinoni et al., 2019; Paul et al., 2013). It is recommended to provide transparency on how the mapping was compiled to support accuracy assessment and, if possible, to obtain reference data (e.g., higher resolution satellite data) to better quantify mapping inaccuracies (Chandler et al., 2018; Smith et al., 2006). In data scarce regions like the Dry Andes of Argentina, reference data is often not available. Thus, transparency about the datasets and software used, the mapping procedure, landform definitions, and associated errors and uncertainties is imperative. To minimize inaccuracies in our inventory, we (1) used satellite imagery from austral summer with minimal snow and cloud cover. (2) Prior to mapping, we established consistent definitions and respective guidelines for mapping block- and talus slopes (see Section 1), and (3) the results of the first two manually mapped key sites were revised in the field (ground truthing). (4) Mapping was performed by one operator and reviewed by three co-authors before entering the predictive models. (5) The combination with the national inventory (IANIGLA-CONICET, 2018) prevents an overestimation of block- and talus slopes at the expense of these cryospheric landforms. Still, mapping uncertainties exist due to the medium resolution of the available data (Google Earth Pro, TanDEM-X DEM and derivatives at 12 m resolution) and the associated size threshold for landforms. Working in areas with limited data availability is intertwined with higher uncertainties, but yields valuable and needed information.

We identified several topoclimatic and geomorphic conditions that favor or inhibit the formation of block- and talus slopes. Conversely, their occurrence provides important insights into their surrounding environment (Brenning, 2005; Messenzehl et al., 2017).

Blockslopes are dominant mesoscale landforms in the periglacial belt and are found primarily at altitudes above 4,000 m in the ANC. They occupy open middle to upper slopes, gently sloping lateral and horizontal ridges, and low-relief plateaus. Slope angles are evenly distributed up to 35°, but remain largely constant along individual slopes. Garleff and Stingl (1983) found an almost uniform distribution of all slope inclinations up to ∼35° on blockslopes in the semi-arid Andes, similar to findings by Clow et al. (2003) in the Colorado Front Range. However, they generally exhibit a narrower range of slope angles around 20°–38°, depending on the angle of repose of the prevailing bedrock (e.g., Selby, 1974; Ballantyne, 2018; French, 2017; Fort and van Vliet-Lanoe, 2007). Unlike talus slopes, blockslopes are covered by finer surface material, with no indications of grain-size induced sorting, and are intersected by small, isolated and resistant bedrock outcrops. They are often bounded downslope by large rock formations, talus slopes, and rock glaciers. Where these landforms do not limit blockslope propagation, they occupy entire slopes down to the valley bottoms. Extensive blockslope coverage down to the basal slope is mostly found on NE-exposed slopes of the ANC, while SW-exposures are dominated by an upslope sequence of talus slopes, bedrock, and blockslopes with occasional bedrock outcrops (see Figure 8). Blockslopes demonstrate a distinct horizontally and vertically elongated, undissected morphometry that has also been described in numerous studies (e.g., French, 2017; Köhler et al., 2024; Garleff and Stingl, 1983; Höllermann, 1983), contributing to their varying terminologies (compare Section 1). Processes such as erosion, accumulation, channeling, and gully formation are mostly absent, as confirmed by field observations and geomorphological mapping. Local incision is limited to snow and glacier meltwater streams, and major erosion and deposition are more likely to be anthropogenically induced (e.g., road construction, see Figure 1C). Their topographic position, shape and characteristics of the contributing area do not promote the production and deposition of allochthonous material (Weiss, 2001). Unlike talus slopes, they are least likely to develop beneath extensive bedrock outcrops with ongoing rockfall activity and steep, laterally converging topography. Blockslopes are described as characteristic features in upland areas with relatively uniform geology and high porosity that have not been affected by glacial steepening (Ballantyne, 2018; Trombotto Liaudat et al., 2014; Selby, 1974). Although the resolution of the available geological data does not allow for more precise statements, this observation aligns with the widespread occurrence of blockslopes in the relatively homogeneous geology of the ANC (see Table 1, Supplementary Figure S5). These observations support their classification as allochthonous debris deposits (Ballantyne, 2018; Höllermann, 1983; van Steijn, 2002; Otto, 2006; French, 2017; Trombotto Liaudat et al., 2014).

The altitude, position-dependent distribution, and topographic openness of blockslopes suggest climatic control factors, e.g., exposure to high solar radiation, large seasonal and diurnal temperature fluctuations, and wind action. Due to its subtropical location, the ANC experiences exceptionally high solar radiation, comparatively low cloud cover throughout the year, and low solar absorption by atmospheric constituents (Schrott, 1998; Viale et al., 2019; Barry and Chorley, 1992). Peaks in global radiation exceeding 1400 W/m² and maximum daily sums of 35.6 MJ/m² were recorded above 4,000 m asl by Schrott (1998) in December 1990 and 1991, which is roughly consistent with more recent measurements by Pitte et al. (2022) (see Section 2). Although aspect and PISR are not the primary drivers of block- and talus slope distribution, their influence is evident in the observed asymmetric zonation, with more extensive blockslope distribution on NE-facing slopes (see Figure 8). The high prevalence of blockslopes on exposed upper slopes, gently inclined ridges, and low-relief plateaus favors high solar irradiation, aridity, and elevated ground temperatures depending on altitude, which also influences the formation of permafrost (Gruber, 2005; Arenson and Jakob, 2010). Decreasing landform probabilities with rising topographic wetness indices (TWI) further indicate that blockslopes form in well-drained, arid areas, although both target landforms are generally associated with dry conditions (low TWI), consistent with the dry climatic conditions in the ANC (see Section 2) (Pitte et al., 2022; Schrott, 1998). In addition, the high exposure of blockslopes to wind action results in minimal snow cover and duration, further promoting dryness and little protection against temperature fluctuations throughout the year (Clow et al., 2003; Köhler et al., 2024). This leads to intense physical weathering under extremely cold and arid conditions, which is strongly associated with blockslope formation (e.g., French, 2017; Fort and van Vliet-Lanoe, 2007; Iwata, 1987; Trombotto Liaudat et al., 2014). Intensive physical rock decomposition, mainly caused by salt weathering and frost action, is considered to be the dominant driver in forming the thin debris layer covering the retreating bedrock they develop on (Trombotto Liaudat et al., 2014; Ballantyne, 2018; Shaw and Healy, 1977; Höllermann, 1983). Brettschneider (1980) recorded the highest weathering intensities on subtropical sun-exposed mountain slopes around 30°, aligning with the maximum blockslope distribution between 30° and 40°S on the Southern Hemisphere (Höllermann, 1983).

Grain size compositions range from mainly silt to medium and coarse sand, with surface clasts primarily consisting of block-sized debris (Otto, 2006; Selby, 1974; Garleff and Stingl, 1983). The thin debris layer varies from decimeters to several meters in thickness (Ballantyne, 2018; Fort and van Vliet-Lanoe, 2007; Garleff and Stingl, 1983). Downslope movement occurs at low velocities through various processes, including solifluction, frost creep, and potentially permafrost creep, depending on the presence of ground ice (Ballantyne, 2018; Selby, 1974; Fort and van Vliet-Lanoe, 2007; Garleff and Stingl, 1983). Denudation by wind action, rather than linear erosion, is responsible for debris removal along slopes (French, 2017; Selby, 1974; Otto, 2006; Höllermann, 1983). Limited downslope accumulation produces a slight basal concavity (slightly positive profile curvature₁₅₁, see Supplementary Figure S3), and may also be a relic of the ANC’s glacial legacy. Still, material supply and removal appear to be in a dynamic equilibrium, resulting in the characteristic rectilinear shape of blockslopes (e.g., Trombotto Liaudat et al., 2014; Ballantyne, 2018; French, 2017; Brenning, 2005). With increasing elevation, unvegetated, rocky blockslopes become predominant in the ANC. However, in climatically suitable areas at lower altitudes, a sparse vegetation composed of adapted species was observed in the field, suggesting an overall lower level of process activity.

Even sparse vegetation cover is almost absent on active talus slopes, indicating much higher process activity on one of the most common debris storage landforms in mountain environments (Messenzehl et al., 2017; Sass, 2006; Phillips et al., 2009; Scapozza et al., 2011). In the Dry Andes above 3,700 m asl, talus slopes extend up to 5,000 m asl, favoring large contributing areas (≥10³ m²) with steep, massive bedrock outcrops. Therefore, they are mostly concentrated on middle to lower slopes below ascending headwalls, cirques, rock glacier rooting zones, or cliff bases. They may extend into the floodplain, where they are partly covered by alluvial fans and eroded by fluvial processes (Alonso and Trombotto Liaudat, 2013; Köhler et al., 2024). Their vast distribution is embedded in a heterogeneous, rugged terrain with juxtaposed small-to large-scale landforms (Messenzehl et al., 2017). Talus slopes are more prevalent on SW-facing lower slopes in the ANC, resulting in lower exposure to solar radiation (see Figures 8, 9). As PISR is ranked with medium to high importance by the RF, talus slope formation is favored by topographically induced shading, leading to cooler conditions year-round (Phillips et al., 2009).

Köhler et al. (2024) identified the characteristics of the contributing area as useful indicators for analyzing and distinguishing the distribution of target landforms, as they relate to the type of material supply (in situ or ex situ), process intensity, and topographic position. This is confirmed by the greater influence of these variables on the prediction of talus slopes, consistent with their formation by gravitational processes from adjacent steep slopes (Brenning, 2005; Moore et al., 2009; Ballantyne, 2018; Alonso and Trombotto Liaudat, 2013). While the probability of blockslopes decreases with increasing mean slope and roughness of the contributing area (MRCA, MSCA), talus slopes become more likely (see Supplementary Figure S3). Since rockfall activity and granular disintegration of cliff faces are the primary sources of talus accumulation, their formation is mainly controlled by weathering and topography, forming steep and thick talus deposits of varying sizes and shapes over time (Brenning, 2005; Ballantyne, 2018; Messenzehl et al., 2017; Clow et al., 2003). Their slope inclination corresponds to the angle of repose of the coarse rockfall material, resulting in a much higher prevalence of talus slopes with angles >35° than blockslopes. Slope gradients between 31° and >40° are frequently reported (Clow et al., 2003; Messenzehl et al., 2017; Scapozza et al., 2011).

The ANC encompasses a wide range of talus slopes that vary in size, shape and thickness depending on the characteristics of the adjacent bedrock, and topography. Talus material may be redistributed by snow avalanches, fluvial processes, and debris- and dry grain flows (Otto, 2006; van Steijn, 2002). Compared to blockslopes, these landforms are significantly smaller and more spatially confined, as reflected in the smaller optimal moving window sizes determined for predictive modeling (Sîrbu et al., 2019). We identified the three main types of talus slopes through geomorphological mapping and field observations: (1) talus sheets with fairly uniform rockfall activity, along with widespread (2) talus cones, and (3) coalescing talus cones, formed by extensive channeling of rockfall debris and lateral merging of individual cones (see Figure 1D) (Ballantyne, 2018; Brenning, 2005). The accumulation and channelization of material favor the formation of talus sheets and cones in the ANC, supported by the greater influence of horizontal (profile) and vertical (planform) curvature in the predictive modeling of talus slopes than in blockslopes. In addition to their predominantly straight shape, talus slopes exhibit a slight tendency toward upward convexity and lateral concavity, indicating horizontal material deposition and vertical convergence of flow (see Figure 9 and Supplementary Figure S3) (Etzelmüller et al., 2001; Sîrbu et al., 2019; Deluigi et al., 2017). Although talus slopes also have low topographic wetness indices, the conditions for moisture concentration are more favorable. Their environmental characteristics promote higher water availability and moisture storage capacity due to reduced exposure to wind and sun. Local snow accumulations are often found at upper talus slopes, especially below large rock cliffs (Brenning, 2005). During field work, we frequently observed erosion rills, debris flows, and alluvial channels likely attributed to meltwater on the surfaces of talus slopes, which could not be captured in the models due to the coarse resolution (Brenning, 2005). The cold climate of the ANC promotes strong physical weathering with frost action and frequent freeze-thaw cycles resulting in bedrock fracturing, and rockfall activity, thus acting as a sediment source (Clow et al., 2003; Messenzehl et al., 2017; Matsuoka and Murton, 2008).

Field observations confirmed a much coarser, poorly sorted debris composition than on blockslopes. Subsurface investigations reveal layers of coarse, open-work material with a higher proportion of decimeter-to meter-sized boulders at the surface and basal talus overlying interbedded fine and coarse-grained clastic material (Ballantyne, 2018; Sass, 2006; Clow et al., 2003; Lambiel and Pieracci, 2008; Scapozza et al., 2011). The deposition of finer material mainly covering silt, sand, and gravel implies that talus formation is not only driven by rock fall deposition but also by alluvial deposition from disintegrated material from source rockwalls (Ballantyne, 2018; Clow et al., 2003). Studies from the Alps and Svalbard using geophysical methods found sediment thicknesses of 3–45 m overlying bedrock (e.g., Hoffmann and Schrott, 2003; Sass, 2006; Clow et al., 2003; Lambiel and Pieracci, 2008). In periglacial environments, talus deposits often contain varying amounts of ground ice and their formation is frequently modified by processes such as periglacial creep, snow avalanches, and debris flows (Ballantyne, 2018; Scapozza et al., 2011; 2015; Sass, 2006; Hilbich et al., 2022). In such environments with ongoing creep, talus slopes may evolve into protalus ramparts and rock glaciers (Otto, 2006; Ballantyne, 2018; Buckel et al., 2021; Scapozza et al., 2011). Their dominant formation at the foot of rockwalls on glacially oversteepened valley sides and cirques in the ANC is conditioned by the glacial heritage that enhances talus supply (Brenning, 2005; Messenzehl et al., 2017). Despite missing information on past or present glacier retreat, increased paraglacial rockfall activity on talus slopes following glacier recession is expected in the ANC (Brenning, 2005).

Using geomorphological mapping and predictive modeling, we significantly reduced the spatial discrepancy between the cryospheric landforms recorded in the national inventory (∼27 km²) and the area potentially underlain by permafrost (regional lower permafrost limit ≥3,700 m asl, ∼686 km² (Trombotto, 2000; Esper Angillieri, 2009); PZI ∼668 km² (Gruber, 2012)). Approximately 516 km² (>75%) of the area is covered by block- and talus slopes, making them extremely abundant in this part of the Dry Andes.

Rock glaciers are the most prominent permafrost landforms in the periglacial belt of the Dry Andes, and their formation and topoclimatic conditions result in a higher permafrost content and thus greater hydrological significance than other periglacial landforms (Corte, 1978; Halla et al., 2020; Arenson and Jakob, 2010; Jones et al., 2019; Schaffer et al., 2019). Beyond these, however, knowledge about permafrost distribution and fresh water storages in periglacial landforms remain limited, despite growing evidence of ground ice presence in areas without rock glaciers (e.g., Hilbich et al., 2022; Mathys et al., 2022; García et al., 2017; Arenson and Jakob, 2010). For the Dry Andes of Argentina, the PZI (Gruber, 2012; see Figure 1, Supplementary Figure S4) and the regional permafrost distribution model developed by Tapia Baldis and Trombotto Liaudat (2020) show that permafrost is not restricted to rock glaciers. Low MAAT at high altitudes, strong temperature gradients, and the rugged mountain topography with varying exposures promote frost action, mechanical weathering, and associated geomorphological processes that define the periglacial domain (French, 2017; Trombotto, 2000). Garleff and Stingl (1983) describe the intensive weathering with high fine material production as a characteristic feature of the semi-arid periglacial stage that is further enhanced by pronounced temperature fluctuations and frequent frost cycles. In addition, widespread and intense rockfall activity produces large accumulations of coarse debris (Lauro et al., 2017; Halla et al., 2020). Depending on the thickness and physical properties of the debris layer, these accumulations effectively insulate the ground and thus provide favorable conditions for ground ice formation and preservation (Arenson et al., 2022). High insolation and wind action, low cloud cover, relative humidity, and precipitation contribute to high sublimation rates and a thin, short-lived seasonal snow cover (Pitte et al., 2022; Ayala et al., 2023; IANIGLA-CONICET, 2018; Brenning, 2005). Minimal snow cover offers fragmented protection against rising summer temperatures and radiation exposure, especially on exposed blockslopes (Arenson et al., 2022; Delaloye et al., 2003; Selby, 1974; Morard et al., 2010). Simultaneously, these conditions reduce the diurnal to seasonal effect of snow insulation and allow deep penetration and circulation of cold air, depending on the physical properties of the debris layer, in winter (Phillips et al., 2009; Lambiel and Pieracci, 2008; Morard et al., 2010). This distinctive thermal regime favors permafrost conditions and ground ice formation, underscoring the high potential of yet unidentified freshwater resources in the periglacial belt of the Dry Andes. Based on our findings, we can evaluate the conditions for the formation and preservation of permafrost based on the environmental distribution conditions of block- and talus slopes.

Overall, the probability and extent of permafrost are significantly higher in talus slopes than in blockslopes due to their differences in spatial distribution and formation conditions. The topoclimatic and geomorphologic characteristics of talus slopes result in thick, coarse debris deposits, greater topographic shading, snow accumulation, and enhanced moisture availability, creating much more favorable conditions for permafrost occurrence. The coarse, open-work debris accumulations lead to high air content near the surface, facilitating deep penetration and circulation of cold air within the talus deposits (Scapozza et al., 2015; Hilbich et al., 2022). Studies investigating the internal structure of talus slopes in the European Alps have revealed variable ice content of up to 60%, primarily located in the lower taluses (Scapozza et al., 2015; Lambiel and Pieracci, 2008; Sass, 2006). Furthermore, talus slopes form geomorphological niches similar to rock glaciers, frequently covering their rooting zones and thus occurring in close proximity to common permafrost indicators (Brenning, 2005). As the formation of protalus ramparts is also an indicator of ground ice (e.g., García et al., 2017; Lambiel and Pieracci, 2008; Ballantyne, 2018), we further mapped their distribution in the upper ANC and SLC, where they were found on at least 1.85% of the surveyed talus slope area, primarily in the lower talus (see Figure 1D).

In contrast, no clearly visible indications of ground ice were identified on blockslopes, although creeping structures were observed on some of these landforms. These resemble small, transversely elongated lobes formed by gelifluction described by García et al. (2017), which are produced by freeze-thaw cycles in the active layer. The authors describe the widespread occurrence of these landforms (referred to as gelifluction taluses in the publication), which are almost exclusively associated with permafrost in the Atacama region of Chile and the arid Andes of San Juan. Yet, the greater aridity caused by high solar and wind exposure in upper, topographically open positions, combined with the thin, in-situ weathered debris mantle, limits their storage capacity and insulation of potential ground ice content (Köhler et al., 2024). However, no studies have examined their internal structure with respect to potential ground ice content, even though blockslopes have been repeatedly described as characteristic periglacial features in the Andes (e.g., Garleff and Stingl, 1983; Hilbich et al., 2022; Brenning, 2005; Köhler et al., 2024) and other cold-climate regions (e.g., Oliva et al., 2022; Ballantyne, 2018; Clow et al., 2003; French, 2017; French and Guglielmin, 1999). While ice content in blockslopes may be limited to small ice lenses or absent depending on local site conditions, they meet several criteria for permafrost presence and their high prevalence in the periglacial belt of the Dry Andes makes them an important subject for further investigation.

Accurately estimating the permafrost content in block- and talus slopes remains challenging due to a lack of subsurface investigations involving boreholes, ground temperature measurements, and geophysical surveys. By modeling their distribution characteristics and analyzing their main controlling factors, however, we identified multiple conditions favorable for permafrost presence in these landforms. Although their permafrost content is likely lower than that of rock glaciers, these periglacial landforms gain hydrological significance due to their widespread occurrence. Estimating the distribution and hydrological significance of ground ice based solely on rock glacier occurrence will inevitably lead to inaccurate quantifications of freshwater resources stored in ground ice and runoff contributions from the alpine periglacial belt (Mathys et al., 2022; Hilbich et al., 2022). A comprehensive asssessment of the periglacial belt in the mountain cryosphere is the basis for further research and accurate quantification.