The Role of Collecting Data on Various Site Conditions Through Satellite Remote Sensing Technology and Field Surveys in Predicting the Landslide Travel Distance: A Case Study of the 2022 Petrópolis Disaster in Brazil

Highlights

What are the main findings?

• To predict the landslide travel distance, it is necessary to consider not only the landslide scale itself, but also the site conditions of the initiation, runout, and deposition zones.
• A substantial proportion of the site condition data was obtained through satellite remote sensing, while a smaller proportion was acquired through field surveys.

What is the implication of the main finding?

• Even in regions with limited data, satellite remote sensing technologies can significantly improve the prediction of the distance reached by landslides.
• The approach provides a scalable, cost-effective, and rapid framework, making it especially valuable for disaster risk reduction in resource-constrained regions.

Abstract

Landslide runout distance is governed not only by collapsed magnitude but also by site-specific geoenvironmental conditions. While remote sensing techniques has advanced landslide susceptibility mapping, its application to runout modeling remains limited. This study examined the role of collecting data on various site conditions through remote sensing and field surveys datasets in predicting the landslide travel distance from the 2022 disaster in Petrópolis, Rio de Janeiro. A total of 218 multivariate linear regression models were developed using morphometric, remote sensing, and field survey variables collected across collapse, transport, and deposition zones. Results show that predictive accuracy was limited when based solely on landslide scale (R² = 0.06–0.10) but improved substantially with the inclusion of site condition data across collapse, transport, and deposition zones (R² = 0.49–0.51). Additionally, model performance was strongly influenced by runout path typology, with channelized flows producing the most stable and accurate predictions (R² = 0.73–0.90), while obstructed and open-slope paths performed worse (R² = 0.39–0.61). These findings demonstrate that empirical models integrating multizonal site-condition data and runout path typology offer a scalable, reproducible framework for landslide hazard mapping in data-scarce, complex mountainous urban environments.

1. Introduction

Landslides are formally classified as geohazards in international frameworks and represent one of the most pervasive and destructive hazards worldwide, being strongly associated with hydrometeorological triggers, particularly intense rainfall, as well as seismic activity and anthropogenic disturbances [1,2]. To mitigate potential landslide impacts, hazard mapping is a fundamental activity. Hazard maps provide essential information for land use planning, early-warning, and evacuation systems, thereby supporting strategies to reduce potential losses and fatalities [3,4,5]. So, numerous studies have been conducted to create more accurate landslide disaster hazard maps [6,7,8,9,10,11,12].

Hazard maps of landslide disasters commonly consist of two parts: initiation of the landslide and its runout and deposition. Several studies have been conducted to assess the spatial patterns of landslide susceptibility, with a focus on landslide initiation [13,14,15]. While, landslide runout, defined as the total horizontal distance traveled by sediments, has been also studied in the last several decades [16,17,18,19,20,21,22]. Within the scope of landslide hazard assessment, predicting the maximum runout distance (or travel distance) is essential for delineating impact zones, informing urban planning, and designing early warning systems [2,16,21]. These methods are broadly categorized into empirical-statistical models and physically-based (or dynamic) simulations.

There are two complementary approaches that can be considered to improve the accuracy of landslide extent prediction. The first approach involves refining the model structure and modelling techniques, while the second approach involves improving the accuracy and quality of the input data used in the model. Recent advancements in remote sensing technology have made it possible to obtain more accurate and diverse spatial geoenvironmental data. Specifically, accurately selecting and collecting the variables that influence landslide extent can improve prediction accuracy. However, many regions still lack sufficient data on site conditions and landslides. In this study, we focused on the second approach. We would therefore like to discuss data collection methods, particularly those that can be used effectively in areas where data is currently insufficient.

Various empirical models focused on correlations between landslide magnitude (i.e., area and volume) and mobility coefficient (i.e., friction coefficient (H/L) or the angle of reach [21,22,23,24,25,26,27,28,29,30,31]. Previous studies showed that runout is a highly non-linear and multifactorial phenomenon, influenced not only by the volume and geometry of the initial failure but also by a complex interplay of terrain attributes, soil and lithological properties, hydrological conditions, and land use [29,32,33]. That is, despite similar landslide magnitudes, the mobility coefficients were widely scattered, suggesting that other factors, such as the landslide materials, water content, and the topography of the runout area, also impacted the landslide runout [7,16,34]. In physically based models, these additional factors are incorporated and considered to control landslide runout [35,36,37,38,39]. There are physically based models that commonly reproduce the actual travel of landslides [40,41,42]. However, it is essential to note that collecting sufficient input data, such as topography, the location of artificial objects, soil properties, and soil water content, for many sites with high landslide susceptibility often requires considerable costs and manpower [20,28].

In view of the aforementioned points, it is considered necessary to gather information on the various site conditions, since the runout distance of a landslide is influenced not only by the scale of the landslide itself, but also by various characteristics of the landslide materials and the site conditions. Advances in remote sensing technology have recently facilitated the acquisition of various spatial data, including topography, vegetation, and artificial objects [33,43,44,45,46]. Furthermore, recent studies have confirmed that even in areas where comprehensive landslide information is conventionally lacking, satellite imagery will provide important information for understanding the characteristics of landslides [47,48,49,50,51,52]. However, despite the existence of numerous studies that have utilized satellite imagery to assess spatial patterns of landslide susceptibility, i.e., initiation zone, there is a paucity of research that has focused on evaluating the runout of landslides [11,53]. Therefore, this study aims to advance landslide runout modeling by developing and evaluating a set of statistical-empirical models based on geoenvironmental site conditions derived from both remote sensing and field survey data. Therefore, the primary objective of this study was to evaluate the effectiveness of remote sensing-derived site-condition variables in predicting landslide runout distances in tropical densely urbanized mountainous regions. Specifically, we aimed to address the following research questions: (i) to what extent can the accuracy of landslide runout distance predictions be improved by incorporating the spatial distribution of site-condition variables obtained from satellite remote sensing data? (ii) in which landslide process zones—initiation, transport, or deposition—is the use of site-condition information most effective in enhancing runout prediction? To explore these questions, we applied a multivariate linear regression framework using the simplest statistical approach as a first step toward assessing the predictive utility of remote sensing variables. The study focused on the slopes surrounding the base of landslides in Petrópolis, Brazil, an area that has experienced severe damage from landslides runout in recent years.

2. Study Area

2.1. Petrópolis Municipality

Petrópolis is a historical and touristic imperial city in the mountainous Rio de Janeiro state region in southeastern Brazil (Figure 1). It covers a total area of 791,144 km² and has a population of 278,881 [54]. Initially dominated by the Atlantic Rainforest biome, the landscape has undergone significant anthropogenic changes, creating a mosaic of urban areas, croplands, and pastures [55,56]. These alterations have increased the city’s vulnerability to natural hazards, with over 15,000 households in high or very high-risk zones [57].

Petrópolis is primarily located within the geomorphological context of the Serra do Mar Escarpment, with elevations ranging from 40 to 2500 m.a.s.l. and an average elevation of 845.5 m. The city features high plateaus, mountainous and high montane regions, dissected and gentle hills, steep mountainous escarpments, and areas with colluvial-talus and alluvial-colluvial deposits [58]. The regional geological setting refers to the Ribeira Mobile Belt, predominantly composed of highly metamorphic rocks such as gneisses with well-developed foliation and syn-tectonic granitoid rocks [59]. Locally, the city is underlain by the Neoproterozoic Rio Negro Complex, characterized mainly by migmatites and granitoids [60]. The soils predominantly consist of cambisol, latosol, and neosol [61]. These soils are primarily characterized by low-fertility clays, good drainage, and high aluminum saturation [62,63]. The steep slopes in the region inhibit the formation of deep soil layers, resulting in relatively shallow soils overlying the bedrock [56], evidenced by the frequent presence of rocky outcrops [64]. The climate is tropical highland type, marked by mild yearly temperatures and well-distributed rainfall. The precipitation is approximately 2200 mm [56]. The stormy season extends from October to March, with monthly rainfall often exceeding 230 mm/month [65].

2.2. Catastrophic Sediment Disasters in Petrópolis

The earliest recorded sediment disasters in Petrópolis occurred in January 1895, resulting in one fatality [66]. Notable events include the March 1966 disaster, where landslides, debris flows, and floods due to rainfall (320 mm over 18 h; and 675 mm over 3 days) caused 80 deaths and displaced 400 people [67]. In March 1979, another event resulted in 87 fatalities and left 300 homeless due to landslides [66,68]. On 5 February 1988, a substantial quantity of cumulative rainfall (776 mm over 24 days) led to a severe disaster that partially destroyed the city, resulting in over 134 fatalities and leaving more than 1000 people homeless [69]. In December 2001, an intense rainfall (300 mm over 14 h) triggered another episode, causing 50 fatalities and displacing 469 people [66]. The January 2011 disaster is considered the most significant in Brazilian history and the eighth worst globally [70,71]. Heavy rainfall (573.6 mm over 10 days; 260 mm over 24 h) triggered 3600 landslides, resulting in 947 deaths, over 400 missing persons, and more than 50,000 homeless. These numbers represent the cumulative losses across the seven severely affected municipalities, including Petrópolis [72]. In March 2013, intense rainfall (390 mm over 24 h) resulted in 34 fatalities, and in November 2016, another episode (155 mm over 24 h) caused two deaths due to rockfalls [66]. In 2022, the city experienced two torrential rainfall episodes, 15 February (258 mm/3 h) and 20 March (534 mm over 24 h), triggering several landslides (Figure 1), leading to 231 fatalities [73], intense damage to urban infrastructure, commercial and industrial areas. In March 2024, heavy rainfall (300 mm over 24 h) triggered 580 landslides and flooding, resulting in nine deaths and leaving 600 people homeless [74].

3. Materials and Methods

The February 2022 disaster was selected for detailed analysis because it represents the most tragic sediment-related event ever recorded in the municipality of Petrópolis, remains underexplored in the scientific literature, and was triggered by an extreme rainfall intensity of 258 mm in just 3 h.

3.1. Data Collection and Production

3.1.1. Landslide Inventory

The landslide inventory used in this study, comprising 171 mapped landslides, was derived from a pixel-based change detection approach using satellite imagery from the Landsat 8–9 Operational Land Imager (OLI), with a spatial resolution of 30 m. Pre- and post-event images were acquired from the United States Geological Survey (USGS) Earth Explorer platform [75], corresponding to dates before the landslide event on 20 March 2022, on 31 January 2022 (summer), and another after the event on 13 April 2022 (autumn). Landslide-affected areas were identified through a pixel-based change detection approach applied to spectral variations and subsequently converted into vector polygons for spatial analysis. To address geometric distortions inherent to pixel-based outputs, all polygons were manually validated and refined using satellite imagery in Google Earth Pro, ensuring alignment with actual geomorphic features. The landslide affected area were subdivided into two, collapse (C) and transport-deposition (TD) zones (Figure 2) [76,77]. Landslide area was calculated in a Geographic Information System (GIS) by delineating polygons and computing their surface area based on projected geometry. Landslide volume were estimated using geometric approximations. In the absence of pre- and post-event digital elevation models, it was not possible to calculate volumes through differential topography [2,20]. Therefore, we employed an empirical approach in which volume was approximated by multiplying the affected area by an assumed average thickness [29,78]. Following common practice in the literature and specific analyses of the 2022 disaster [64,72,73,79,80], an initial fixed thickness of 1 m was adopted. To evaluate the uncertainty associated with this assumption, a sensitivity analysis was performed by recalculating volumes with alternative thickness values of 0.5 m and 1.5 m. The results indicated that statistical performance remained unchanged across different thickness values.

To assess landslide mobility, we employed morphometric indices based on the relationship between the elevation drop and the runout distance of the landslide mass (Figure 2). For clarity, notations referring to totals are denoted with the subscript t—Ht (total height), Lt (total length), Vt (total volume); collapse-specific metrics with the subscript c—Hc, Lc, Wc, Ac, Vc; and transport–deposit metrics with the subscript td—Htd, Wtd. The first parameter used was the mobility index, commonly expressed as the ratio between the total fall height (H) and the horizontal travel distance (L), denoted as H/L. This ratio serves as a proxy for the apparent coefficient of friction, reflecting the degree of energy dissipation during movement [22,23]. Additionally, we calculated the equivalent of friction angle (α) or angle of reach, which is derived from the inverse tangent of the H/L ratio: tan(α) = H/L. This metric provides a physically interpretable measure that approximates the average slope of the landslide trajectory and has been widely adopted as a standardized mobility index in landslide studies [21,29]. The angle of reach enables the comparison of mobility across different landslide types and volumes, assuming that more mobile landslides travel longer distances relative to their fall height, thereby yielding lower α values. The authors demonstrated that while smaller landslides tend to exhibit higher angles of reach due to greater influence of local terrain and frictional resistance, large-volume landslides typically show a flattening trend in α values, attributed to momentum-driven dynamics and reduced relative friction. Therefore, the H/L ratio and angle of reach (α) serve as effective, scale-sensitive indicators of landslide mobility, particularly when rheological data or dynamic simulations are unavailable. To provide a detailed analysis of the influence of on-site condition data on landslide runout behavior and reduce the variability, the landslide inventory was categorized into four mobility classes based on mobility ratio (H/L): low mobility (1.00 ≤ H/L ≤ 3.00), moderate mobility (0.70 ≤ H/L < 1.00), high mobility (0.30 ≤ H/L < 0.70), and very high mobility (0.07 ≤ H/L < 0.30) [21].

3.1.2. Site-Condition Datasets

-

Dataset Source and Resolution

The selection of relevant landslide conditioning factors is a critical step in accurately modeling landslide behavior, especially in densely populated regions [1]. Guided by the understanding of landslide processes and the availability of spatial data [19,81], this study selected a total of 23 conditioning factors, which were systematically organized into four geoenvironmental thematic categories: (i) quantitative remote sensing—RS_Qt, (ii) qualitative remote sensing—RS_QL, (iii) quantitative field survey-based data—FS_QT, and (iv) qualitative field survey-based data—FS_QL (

Table 1. Characteristics of site-condition thematic datasets.

Category	Type	Variable	Source
Remote Sensing (RS)	Quantitative (Qt)	Altitude (ALT)	DEM SRTM (30 × 30 m)
		Slope angle (SLP)
		Aspect (ASP)
		Plan curvature (PLCV)
		Profile curvature (PRCV)
		Topographical Position Index (TPI)
		Topographic roughness index (TRI)
		Length-Slope Factor (LSF)
		Stream Power Index (SPI)
		Sediment Transport Index (STI)
		Topographic Wetness Index (TWI)
		Drainage density index (DDI)
		Normalized Difference Water Index (NDWI)	Landsat 8 OLI (30 × 30 m)
		Normalized Difference Vegetation Index (NDVI)
		Normalized Difference Building Index (NDBI)
		Road density index (RDI)	[84]—1:25,000
	Qualitative (Ql)	Geomorphological units (GeoU)	[58]
		Land Use Land Cover (LULC)	[55]
Field Survey (FS)	Quantitative (Qt)	Lineament—Fault density index (LFDI)	Geological maps
		Structural density index (SDI)	1:100,000
		Rainfall (RNF)	1:50,000
	Qualitative (Ql)	Lithology type	Geological map (1:100,000)
		Soil type	Pedological map (1:100,000)

). The remote sensing dataset included a digital elevation model (DEM) derived from the Shuttle Radar Topography Mission (SRTM) (30 × 30 m resolution) [82]. Additionally, orthophotos at a scale of 1:30,000 [83] were employed as high-resolution visual references to support the interpretation of topographic features, infrastructure, and geomorphological context. Satellite imagery from Landsat 8 (Section 3.1) was also utilized to generate spectral indices relevant to terrain and surface condition analysis. In this study, we investigated the availability of Sentinel-2 imagery with high spatial resolution. However, the available scenes from the days immediately following the disaster in 2022 showed extensive cloud cover in the region of interest. As emergency and reconstruction activities in the damaged area significantly altered the site conditions after 10–15 days, Landsat OLI imagery was prioritized as it provided the best available coverage during the critical time window. Land use and land cover (LULC) data from the MapBiomas Collection 8 (1985–2022) were integrated into the analysis to characterize the anthropogenic and environmental settings that influence landslide behavior. This dataset, derived from Sentinel-2 imagery at a 10 × 10 m resolution [55], enables the detailed classification of urban areas, forest formations, pasturelands, and mosaic land use patterns, providing essential context for interpreting human-induced impacts on landslide extent. For this study, we employed a subset of classes restricted to those present within the landslide-affected urban area of Petrópolis, ensuring that the analysis captured only the local environmental conditions relevant to the event. To ensure consistency, the LULC layer (10 m) was resampled to a common 30 m grid prior to incorporation into the landslide inventory dataset. This harmonization minimized spatial misalignment and ensured comparability across datasets. Additional spatial layers included geomorphological units (1:80,000 scale), derived from the interpretation of orthophotos at 1:30,000 resolution [58], and road network (1:25,000) data obtained from OpenStreetMap [84]. Field survey thematic maps were incorporated to complement the environmental characterization of the study area. These included geological maps [60,85] and a soil type map [61], both at a 1:100,000 scale, as well as a rainfall isoet map representing average annual precipitation at a 1:50,000 scale [86].

-

Data Production

The geospatial data were processed and analyzed in QGIS and SAGA GIS [87]. The DEM was used to calculate widely adopted topographic parameters for landslide modeling. All terrain derivatives were computed for each 30 × 30 m grid cell, corresponding to the native resolution of the original SRTM dataset (Table 1; Figure 3). Topographic metrics included altitude (ALT), slope angle (SLP), aspect (ASP), plan curvature (PLCV), profile curvature (PRCV), topographic position index (TPI), topographic roughness index (TRI), and the length-slope factor (LSF) [78,81,88,89,90]. We also calculated topographic characteristics which have been used as a proxy of hydrological and sediment transport processes, such as the topographic wetness index (TWI), stream power index (SPI), and sediment transport index (STI) [91].

Spectral images were employed to estimate normalized indices that served as proxies for on-site environmental conditions (Table 1; Figure 3). The Normalized Difference Vegetation Index (NDVI) reflects vegetation health and density. NDVI was calculated based on ((NIR − RED))/((NIR + RED)), for Landsat 8, NIR corresponds to near infrared band 5 and Red to band 4 [92]. Normalized Difference Water Index (NDWI) provides a proxy for surface moisture conditions, with higher values indicating wetter soils or standing water. NDWI was calculated based on ((NIR − SWIR1))/((NIR + SWIR1)), for Landsat 8, NIR corresponds to band 5 and SWIR1 to short-weave infrared band 6 [93,94]. NDBI was calculated based on ((SWIR1 − NIR))/((SWIR1 + NIR)), for Landsat 8, SWIR1 corresponds to band 6 and NIR to band 5 [44]. The normalization spectral indices range is between −1 and +1.

To quantify structural, hydrological, and anthropogenic linear features influencing landslide behavior, four density-based indices—Drainage Density Index (DDI), Lineament–Fault Density Index (LFDI), Structural Density Index (SDI), and Road Density Index (RDI)—were computed using line density analysis (Table 1; Figure 3). Each index was calculated by applying a line density function over vector line datasets within a defined radius (kernel), producing a continuous raster surface that expresses the total length of linear features per unit area (km/km²). The DDI was derived from river and stream networks and reflects terrain dissection and potential soil saturation near channels. The LFDI was calculated from mapped geological faults and lineaments, representing zones of structural weakness [6,95,96]. The SDI combines lithological boundaries and discontinuities to capture broader structural complexity. These three indices were based on geological layer maps [60,85]. The RDI, computed from the OpenStreetMap [84] road network, serves as a proxy for anthropogenic disturbance and infrastructure exposure [95,97]. Density indices dataset was generated with a uniform spatial resolution of 30 m × 30 m. This standardization ensured comparability among indices (RDI, LFDI, SDI, and DDI) and avoided scale-related biases in the subsequent analyses.

Climatic influence was incorporated through long-term precipitation data (Table 1; Figure 3), represented by a 30-year (1977–2006) average annual rainfall [86]. The original map was available as a vector layer containing point data from meteorological stations across the region. To integrate this information into the spatial modeling framework, the point layer was interpolated into a continuous raster surface using the Inverse Distance Weighting (IDW).

-

Data Extraction

To characterize the on-site conditions influencing landslide behavior, we extracted geoenvironmental variables at three reference points for each mapped landslide body: the initiation point of collapse (C), the beginning of the transport phase (T), and the end of deposition (D) (Figure 2). These points were manually delineated based on visual interpretation of orthophotos and satellite imagery. At each point, values from all continuous and categorical raster layers were extracted using the native tools available in QGIS. Specifically, the “Sample Raster Values” tool was used to assign numerical values (e.g., slope, TRI, NDVI) to each point based on the underlying raster cell. In contrast, categorical values (e.g., land use, soil type, lithology) were assigned to each point by spatially intersecting point locations with vector-based polygon layers, using the “Join attributes by location” tool.

Table 2. Description of qualitative site-condition thematic datasets.

Qualitative	Code	Description
Lithology	NP23_gamma_1rn	Neoproterozoic diorites, tonalites, gabbros, and gneisses.
	NP3_gamma_1rnb	Neoproterozoic gneisses and migmatites.
	NP3_gamma_1so	Neoproterozoic granites.
	O1_gamma_6nfan	Phanerozoic granites.
	NP2_gamma_3sor	Neoproterozoic orthogneisses.
Soil	AR3	Rock outcrop, associated with secondary cambisols and tertiary litholic neosols.
	LVAd24	Dystrophic Red-Yellow latosol, clayey or loamy.
	CXbd4	Haplic cambisol Tb, dystrophic, medium or argillic; associated with a talic B horizon.
	CXbd6	Haplic cambisol Tb, dystrophic, clayey or medium; associated with a talic B horizon.
	Urban area	Urban area
Geomorphological unit	Alluvial-colluvial slopes	Gently inclined depositional surfaces (5–10°) composed of poorly sorted sandy-clayey to clayey-sandy colluvial materials, interdigitated with flat fluvial plain deposits.
	Colluvial-talus slopes	Depositional slopes (5–10°) composed of very poorly sorted colluvial materials with sandy-clayey to clayey-sandy matrix and abundant blocks, interdigitated with gently inclined alluvial-colluvial ramp deposits. Typically located at the base of steep slopes and escarpments.
	Dissected hills and low hills	Convex-concave slopes and rounded hilltops with relief amplitude between 50 and 120 m and slopes of 5–20°. Moderate drainage density.
	Flood plains	Sub-horizontal surfaces (0–3°) located in valley bottoms, composed of well-sorted sandy, sandy-clayey, or clayey deposits.
	High hill domains	Convex-concave geometry with well-dissected terrain, rounded to pointed summits, relief amplitude of 80–250 m, slopes of 10–35°, and moderate to high drainage density.
	Mountainous domains	Straight to concave slopes with aligned ridge tops, pointed or slightly rounded, relief amplitude >300 m. High drainage density, with steep slopes (20–45°) and occasional sub-vertical rocky walls.
	Mountainous escarpments	Highly dissected, steep slopes (>300 m relief amplitude, 30–45° slope), with straight to concave profiles, rocky cliffs, and sharp or aligned ridge tops. High drainage density.
Land Use and Land Cover	Forest formation	Vegetation types with a predominance of tree species with high-density continuous canopy.
	Mosaic of uses	Farming areas where it was not possible to distinguish between pasture and agriculture. In Petropolis is often associated with the transition between informal settlements and areas that have been deforested.
	Pasture	Natural or planted, related to farming activity.
	Rock outcrop	Naturally exposed rocks in the terrestrial surface without soil cover, often with partial presence of rock vegetation and high slope.
	Urban area	Urban areas with a predominance of non-vegetated surfaces, including roads, highways and constructions.

provides a summarized description of the qualitative site-condition classification variables considered in this study. This point-based sampling strategy ensured that each landslide was represented by a standardized set of site-specific parameters across its three process zones (C, T, D), enabling consistent and robust statistical comparison and modeling.

3.2. Statistical Analyses

This study conducted the statistical analyses using Jamovi version 2.5.3, an open-source statistical software.

3.2.1. Factor Analyses (FA)

Spearman’s rank correlation coefficient (SRC) test (ρ) was used for nonparametric data distribution to measure the strength and direction of the association between the transport-deposited (TD) length and the site-condition quantitative parameters. The strength of the correlation is interpreted based on: 0–0.2 (very weak), 0.2–0.4 (weak), 0.4–0.6 (moderate), 0.6–0.8 (strong), and 0.8–1 (very strong) [98]. Analysis of Variance (One-way ANOVA) was also used to investigate the effect of qualitative factors, such as soil type, geomorphological unit, and lithological types, on the landslide’s TD length [99]. Due to the non-normal distribution of the data, a non-parametric approach, the Kruskal–Wallis (KW) test, was employed [100,101]. This method assesses whether significant differences exist between the medians of multiple groups across different categorical variables [102]. The KW test provides a chi-square statistic (χ²) and a p-value indicating whether significant differences exist between the groups’ medians. Also, epsilon-squared (ε²) was calculated to estimate the effect size, classifying it as small effect (ε² < 0.01), moderate effect (0.01 ≤ ε² < 0.08), and significant effect (ε² ≥ 0.08) [103].

3.2.2. Linear Regression Models

Multivariate Linear Regression (MLR) is a widely used statistical technique designed to model the relationship between a single dependent variable and multiple independent variables Equation (1) [33,104,105]. model performance was evaluated using the coefficient of determination (R²) and Root Mean Squared Error (RMSE).

$$Y_i={\beta }_0+\left({\sum }_{K=1}^K{\beta }_K{X}_{iK}\right)+{\in }_i,$$

(1)

where Y_i represents the dependent variable for the ith observation, β₀ is the intercept term, and ${\sum }_{K=1}^K{\beta }_K{X}_{iK}$ corresponds to the summation of the product of coefficients β_K and their corresponding independent variables (X_iK) across all predictors (K). The term ∈_i is the error term (residual), accounting for the deviation of observed values from those predicted by the model. The coefficients (β) were estimated using the Ordinary Least Squares (OLS) method [106,107].

3.3. Research Design

To systematically investigate the controlling factors of site condition parameters on landslide runout distance, we adopted a sequential modeling framework (Figure 4). The research design was structured into three progressive phases, each addressing specific analytical objectives: (i) development of standalone models to evaluate the individual effects of distinct variable sets; (ii) construction of combined models to explore the effects of integrating multiple data sources; and (iii) development of multi-stage models to explore complex interactions across landslide process zones. A total of 218 MLR models were developed based on the complete landslide inventory (n = 171), ensuring a robust empirical basis for performance evaluation. The Model 1 comprised univariate models using post-failure morphometric (PFM) parameters from the collapsed zone, such as height (Hc), length (Lc), width (Wc), area (Ac), volume (Vc), and logarithmic values (LogAc and LogVc) (Figure 4). These variables are widely cited in runout prediction literature [2,19,21,29,104]. Model 2 accessed the individual predictive power of geoenvironmental site-condition variables, considering quantitative and qualitative inputs (Table 2). These were organized by source into remote sensing-based (2A) and field survey-based (2B) data acquisition approaches (Figure 4). Model 2 to 6 were independently developed for the collapse, transport, and deposition points. This approach allowed us to assess the specific influence of site-condition variables at each stage of landslide propagation, recognizing that different factors may control runout [18,32,39].

In the second modeling phase, we assessed whether integrating distinct data sources could enhance predictive accuracy [16]. Model 3 were developed by combining variables selected in factor analysis (Section 3.2.1), identified in four groups: remote sensing-based, quantitative (3Aqt), qualitative (3Aql), and field survey-based, quantitative (3Bqt), qualitative (3Bql). To control for redundancy, multicollinearity diagnostics were performed using the Variance Inflation Factor (VIF) and Tolerance (TOL) statistics. VIF values were interpreted using standard thresholds: VIF ≤ 5 indicating low collinearity, 5 < VIF ≤ 10 acceptable, and VIF > 10 as high [12,108]. Tolerance (1-VIF) was used as a complementary indicator, with values above 0.2 generally considered acceptable. Predictors exhibiting high collinearity were excluded to improve model robustness. Model’s predictor significance was further evaluated using the F-statistic, which tests the null hypothesis that all regression coefficients are equal to zero. A significant F-value indicates that the model explains a substantial portion of the variance in the dependent variable [12,109]. In Model 4, we systematically assessed the combined effect of grouped remote sensing, field survey, quantitative and qualitative geoenvironmental site-condition variables developed in model 3. In Model 5, post-failure morphometric parameters were combined with different geoenvironmental datasets (from Model 4). Model 6 further explored the complete interaction between PFM parameters and the integral geoenvironmental models for each landslide stage, offering a more general perspective of landslide behavior. Subsequently, in the third modeling stage, we developed Model 7, which combined PFM predictors with Model 6 variables from multiple landslide zones (i.e., collapse–transport—CD), collapse–deposit—CD, and transport–deposit—TD). Finally, Model 8 extended this approach by simultaneously integrating each morphometric descriptor and geoenvironmental variables from the three landslide stages into a single, thoroughly combined model (CTD-based). This configuration aimed to maximize predictive capacity by accounting for spatial continuity, environmental heterogeneity, and failure geometry in a unified framework. Through this stepwise modeling design, we established a systematic comparison of morphometric, remote sensing, and field-based predictors, identifying the influence of site conditioning factors for estimating landslide runout behavior, particularly in data-scarce and urbanized environments. The statistical outputs of the CTD-based models (type 8)—including multicollinearity diagnostics (Variance Inflation Factor—VIF and Tolerance values), standardized regression coefficients, normality tests, heteroscedasticity tests, and residual analyses—are provided in the Appendix B.

To assess the predictive uncertainty of each model, we calculated upper and lower bounds based on a proportional deviation from the observed values [110]. Specifically, we applied a scaled RMSE, using the expression ±K × RMSE, where K is a user-defined adjustment factor that controls the confidence envelope around the 1:1 reference line. In this study, we adopted a margin of K = ±0.30 [107]. The RMSE used for this calculation was obtained by computing the average RMSE across all models within each modeling group (e.g., Model 1) separately for each landslide stage (e.g., collapse). This approach ensured that the uncertainty bounds were tailored to the predictive behavior of each specific modeling context. These bounds were plotted around the 1:1 line in the predicted versus observed plots to visually delineate regions of overestimation (above the upper bound) and underestimation (below the lower bound).

To further assess model robustness and generalization capacity, we implemented a two-step validation procedure. First, a bootstrap resampling strategy was applied to the full dataset of 171 landslides. In each iteration, a bootstrap sample of size 171 was drawn with replacement for model calibration. Because of this procedure, some cases were sampled multiple times while others were excluded; on average, about 63% of the landslides appeared at least once in the training set, while the remaining ~37% (66 cases) formed the out-of-bag (OOB) set, which served as an independent test set. Second, 10-fold cross-validation was applied within each bootstrap training set to further estimate generalization capacity and reduce overfitting. This procedure was restricted to the final CTD-based models (type 8), which represent the culmination of the modeling framework. Generalization performance was evaluated using the R², RMSE, and mean absolute error (MAE).

4. Results

4.1. General Characteristics of Landslides in the 2022 Petrópolis Event

The analysis of initiation zones revealed that the predominant lithological units associated with the landslides were diorites, tonalites, gabbros, and gneisses, collectively representing 75% of the events and belonging to the same geological formation. Secondary occurrences included migmatites and gneisses (16%) and granites (9%). In terms of geomorphological context, landslides were most frequently observed within the high hills domain (44%), followed by mountainous escarpments (26%) and the broader mountainous domain (22%). Regarding soil types, the majority of landslides occurred in urbanized areas with anthropogenic soil disturbance (35%), with natural soil types such as cambisols (29%) and latosols (22%) also showing notable concentrations. Additionally, approximately 14% of the events were located on steep slopes characterized by shallow soils overlaying exposed bedrock, indicating conditions of limited soil depth and high susceptibility to failure. The distribution analysis of landslide areas revealed that moderate-sized landslides are the most common (60%), followed by small landslides (37%) and medium–large-sized occurrences (4%). The volume distribution exhibited a broad spectrum ranging from very small to extremely large events. Most landslides fall within the moderate (33%), small (32%), and medium-large (23%) classes, spanning from 10³ to 10⁶ m³. Larger landslides, exceeding 10⁶ m³, are less frequent (9%). Based on satellite image interpretation, the TD runout zones were classified into four morphological types: channelized (CHAN), channelized–obstructed (CHOB), obstructed (OBS), and unobstructed (UNOB), according to their flow trajectory and physical confinement. OBS paths were the most common (50%), followed by UNOB (22%), CHOB (20%), and CHAN (8%). Natural obstructions included barriers such as vegetation corridors, terraces, and overlapping landslide deposits, while anthropogenic obstructions were mainly related to roads, buildings, and cut-and-fill platforms. The depositional slope angles corresponding to these runout types ranged from 5° to 49°, with an overall average of 18°. In addition, CHAN landslides exhibited the lowest mean slope angles, while OBS and UNOB runouts were typically associated with steeper depositional zones. This suggests that greater slope inclination may favor more open flow trajectories, while flatter terrain promotes lateral spreading and obstruction by surrounding features.

4.2. Factor Analysis

4.2.1. Correlation Between Post-Failure Morphometric Parameters and Landslide Runout

The TD length was strongly associated with total length (0.89) and volume (0.61) but showed only a moderate correlation with total height (0.32). In the collapsed zone, TD length exhibited weak associations with height (0.24), length (0.22), width (0.19), altitude (0.13), and slope angle (0.12). For the transport-deposited zone, TD length exhibited a moderate link with height (0.32) and a weak correlation with width (0.26), transported altitude (0.05), slope angle (0.12), and deposition slope angle (0.02). A weak negative correlation was observed with deposition altitude (−0.05) (

Table 3. Spearman correlation coefficients (ρ) between key post-failure morphometric landslide attributes and TD length.

Post-Failure Morphometric Variables	Transport-Deposit Length
Collapse height (Hc)	0.24
Collapse width (Wc)	0.19
Collapse length (Lc)	0.22
Collapse altitude (ALTc)	0.13
Collapse slope angle (SLPc)	0.12
Transport-Deposit height (Htd)	0.32
Transport-Deposited width (Wtd)	0.26
Transport altitude (ALTt)	0.05
Transport slope angle (SLPt)	0.12
Deposit altitude (ALTd)	−0.05
Deposit slope angle (SLPd)	0.02
Total height (Ht)	0.32
Total length (Lt)	0.89
Total volume (Vt)	0.62

Note: Cells with p values < 0.05 are highlighted in gray to indicate statistical significance.

).

4.2.2. Correlation Between Site-Condition Features and TD Runout Length

Variable selection for each landslide process stage was primarily based on Spearman’s rank correlation coefficients, with a threshold of ρ ≥ 0.10 (

Table 4. Spearman correlation coefficients (ρ) between quantitative on-site conditions parameters and TD length.

Variables	Collapsed	Transported	Deposited
ALT	0.13	0.06	−0.05
SLP	0.12	0.12	0.02
ASP	0.26	0.28	0.15
PLCV	0.18	0.13	0.11
PRCV	0.13	0.13	0.01
TPI	0.23	0.19	0.12
TRI	0.13	0.13	0.07
SPI	0.10	0.11	0.06
LSF	0.02	0.04	−0.01
STI	−0.02	0.03	−0.01
TWI	−0.18	0.13	−0.08
DDI	0.01	0.01	0.05
NDWI	0.15	0.05	0.02
LFDI	−0.14	−0.14	−0.18
SDI	−0.19	−0.18	−0.18
NDVI	−0.26	−0.11	−0.07
RNF	0.05	0.07	0.09
RDI	0.00	0.02	0.05
NDBI	0.23	0.04	0.05

Note: Cells with p values < 0.05 are highlighted in gray to indicate statistical significance.

). Several variables demonstrated consistent correlations across all stages, including ASP (0.15 ≤ ρ ≤ 0.28), TPI (0.12 ≤ ρ ≤ 0.23), SDI (−0.18 ≤ ρ ≤ −0.19), LFDI (−0.14 ≤ ρ ≤ −0.18), and PLCV (0.11 ≤ ρ ≤ 0.18), highlighting their pervasive influence on landslide dynamics from initiation through deposition. However, due to stage-specific, variable selection was tailored separately for each zone. For the collapsed and transport zones, selected variables included ALT, SLP, ASP, PLCV, PRCV, TPI, TRI, SPI, TWI, NDWI, LFDI, SDI, NDVI, NDBI, and RNF. In contrast, the deposition zone was characterized by a subset comprising ASP, PLCV, TPI, TRI, TWI, LFDI, SDI, NDVI, NDBI, and RNF.

4.2.3. Correlation Between Qualitative Parameters and TD Length

In the collapse zone, the one-way ANOVA showed that soil type (χ² = 9.84; ε² = 0.058) had the strongest effect, followed by LULC (χ² = 3.73; ε² = 0.022), on runout distance (Figure 5). Geomorphological units (χ² = 2.56; ε² = 0.015) and lithology (χ² = 0.30; ε² = 0.002) showed a weak correlation. In the transport phase, soil type (χ² = 12.009; ε² = 0.071) and lithology (χ² = 9.94; ε² = 0.058) showed a substantial impact. Geomorphological unit (χ² = 2.44; ε² = 0.014) and LULC (χ² = 2.18; ε² = 0.013) were weakly related with TD length. In the deposition stage, lithology (χ² = 19.03; ε² = 0.112), soil type (χ² = 10.92; ε² = 0.064) and geomorphological units demonstrated significant influence on the TD length (χ² = 10.62; ε² = 0.062). For LULC, results showed a lower influence (χ² = 4.74; ε² = 0.028).

4.3. Linear Regression Models

4.3.1. Univariate Regression Models (1 to 2)

Detailed outputs of the intermediate models (Models 2–6) are provided in Appendix A. The univariate models based on post-failure morphometric parameters from the collapse zone (Model 1) yielded coefficients of determination (R²) ranging from 0.06 to 0.10 (Figure 6). The best-performing variables were the length and LogAc (R² = 0.10). Overall, the results indicate that individual morphometric parameters have limited predictive performance in explaining runout distance. Similarly, models utilizing a solitary site-condition variable (Model 2) exhibited very low predictive power (R² < 0.10).

In summary, the analyses of intermediate models (Models 3 to 6) showed a pattern of progressive improvement in the R² value as a number of parameters used in models increased from one to many (Appendix A). The results from Model 3 showed a slight improvement in performance by using multiple parameters relating to site conditions belonging to a single data group (quantitative or qualitative data from remote sensing or field surveys). In the collapse zone and transport zone, the highest performance was achieved when using a group of quantitative data from remote sensing (R² = 0.14 and 0.11). The model that used data from multiple data groups (Model 4) demonstrated improved accuracy. The highest accuracy was achieved when using data from all field conditions, obtained through remote sensing and field surveys, with an R² value of 0.21.

The addition of landslide morphometric parameters (Model 5) marked a moderate improvement compared with Model 3. The highest accuracy was achieved using collapse length (Lc) and environmental descriptors in the collapse and transport zones (R² = 0.20). In terms of data source, the quantitative variables obtained from remote sensing yielded the strongest performance in the collapse and transport zones (R² = 0.16–0.20). The model that used data from multiple data groups for site condition variables and landslide morphometric parameter (Model 6) reinforced and the highest R² was 0.28 (Appendix A).

4.3.2. Multi-Stage Linear Regression Models (7 to 8)

The multistage integration models (Model 7), combine variables from two landslide stages. In the CT models, the highest R² was achieved by combining collapse length with environmental variables (R² = 0.35; Figure 7), followed by log-transformed variables and height (R² = 0.32). In the CD models, predictive power increased substantially, with collapse length yielding the highest outcome (R² = 0.44; Figure 8), while all other variables ranged between R² = 0.42 and 0.43. A similar pattern was observed in the TD models, where length again achieved the best performance (R² = 0.44; Figure 9). These results confirm that integrating environmental descriptors from multiple process zones enhances model accuracy, with collapse length (Lc) consistently emerging as the strongest post-failure predictor. Among the multi-stage configurations, CD and TD models exhibited equally high and stable predictive performance, with all R² values concentrated between 0.42 and 0.44.

The final multistage morphometric models (Model 8) integrate collapse, transport, and deposit zones (CTD-based) variables, achieving the highest predictive accuracy among all tested configurations (Figure 10). The combination of collapse length (Lc_CTD) with geoenvironmental descriptors yielded the best performance (R² = 0.51), followed closely by logarithmic transformations of volume and area (R² = 0.50). Other metrics such as height, volume, and area also showed consistent results (R² = 0.49), reinforcing the general utility of collapse morphometry across multiple stages. Despite slight differences among variables, all morphometric predictors produced stable and convergent performance in the full multistage context, with narrow variation in R² (0.49–0.51). To verify the stability of these predictors, multicollinearity was evaluated VIF and TOL values. All predictors in the final models presented VIF < 4 and Tolerance > 0.25, indicating no critical multicollinearity. The complete statistical analysis is provided in the Appendix B.

4.3.3. Model’s Type 8 Cross-Validation

The performance of the CTD-based models (type 8) under 10-fold cross-validation is presented in

Table 5. Generalization performance metrics of CTD-based models (type 8) under 10-fold cross-validation.

CTD_Based	10-Fold Cross-Validation
Models	R2	RMSE	MAE
CTD-based	0.12	71.66	48.18
Hc_CTD	0.12	71.63	48.19
LC_CTD	0.15	70.33	47.58
Ac_CTD	0.10	72.56	48.48
Vc1.0_CTD	0.10	72.58	48.79
LogAc_CTD	0.11	71.82	48.06
LogVc1.0_CTD	0.12	71.56	47.76

. Across all specifications, predictive capacity declined sharply compared to training evaluation, with R² values dropping to 0.10–0.15 and error metrics increasing to RMSE ≈ 70–73 m and MAE ≈ 47–49 m. Among the tested models, LC_CTD yielded the best performance (R² = 0.15, RMSE = 70.33 m, MAE = 47.58 m).

5. Discussion

5.1. Can the Landslide Travel Distance Predictions Be Improved by Using Information Obtained from Satellite Remote Senszing?

The results demonstrate that predictive accuracy varied substantially depending on the type and integration of input data (Figure 11a). Models incorporating post-failure morphometric (PFM) parameters achieved lower performance (median R² ≈ 0.08). Also, models based exclusively on remote sensing (RS) or field survey (FS) variables presented consistently low median R² values (below 0.03–0.04). These revealed their limited individual capacity to capture the complexity of landslide runout dynamics in the study area. This relatively modest predictive value of PFM parameters aligns with prior empirical findings, which emphasize that PFM alone cannot reliably predict runout due to high dispersion in mobility coefficients across events of similar magnitude [26,27]. Improvements in model performance were evident when RS variables were combined with PFM datasets (median R² = 0.14). This reflects the additive value of RS-based descriptors. Although RS data are frequently used in susceptibility modeling [111,112], their application in runout prediction remains relatively uncommon, particularly when decoupled from physically-based modeling frameworks [11]. The present results show that such variables can support empirical modeling if used in conjunction with site-specific parameters. The highest performance was obtained when RS, FS, and PFM variables were integrated in a unified model structure (median R² > 0.35). The integrated configuration outperformed all others by a substantial margin, indicating that no single data type is sufficient to explain runout behavior under the topographic and anthropogenic complexity observed in urban settings.

5.2. Influence of Landslide Stage on Predictive Performance

Model performance increased progressively with the integration of multiple landslide stages, underscoring the cumulative nature of runout dynamics (Figure 11b). Substantial improvements were observed when two zones were combined. Models integrating collapse and transport reached a median R² of 0.32, whereas combinations involving collapse and deposit or transport and deposit achieved higher accuracy (median R² = 0.43). These findings align with conceptual models of landslide propagation, which emphasize that runout results from a sequence of interdependent processes—including initial failure, flow propagation, and terminal deposition—each governed by distinct physical controls [2,20,32]. The enhanced performance of models incorporating depositional components highlights the critical role of terminal-phase conditions. This is consistent with prior findings that emphasize the influence of terrain resistance and stratigraphic contrasts during the final stage of landslide mobility [10,39,78,113]. The highest model performance was observed when all three process stages were integrated (CTD), with a median R² of 0.50. This supports the hypothesis that multistage modeling provides a more comprehensive representation of the physical controls governing landslide mobility, particularly in heterogeneous or anthropogenically modified terrain [7,114].

5.3. Influence of Landslide Runout Path

Hungr (2007) emphasized that confined channels promote compression of the debris mass, increasing basal pore pressure and reducing internal energy dissipation [115]. While McDougall & Hungr (2005) demonstrated that in unconfined settings, debris flows are more susceptible to lateral spreading, velocity decay, and irregular deposition, especially on convex or fragmented surfaces [18]. Studies by Scheevel (2017) and Guo et al. (2020) demonstrated that even minor obstructions can cause redirection, deceleration, or bifurcation of debris flows, depending on their relative position and physical configuration [90,114]. Frodella et al. (2022) documented that built environments often act as physical filters, interrupting the spatial continuity of flows and forcing premature deposition [10]. These previous studies suggested that the flow path geometry and the obstacles in flow path may give large impacts on travel distance of landslide.

Figure 12 shows the best advanced multivariate regression models (Model 6–8), which incorporate collapse length (Lc) along with geoenvironmental variables, stratified by typological TD runout classification (Section 4.1), as channelized (14 landslides), channelized-obstructed (34 landslides), obstructed (86 landslides), and unobstructed (38 landslides), offering a process-based framework to assess how topographic confinement and physical barriers condition the mobility of landslides in complex urban-mountainous settings. It is unfortunate that the sample size is limited, and the results are thought to be subject to considerable uncertainty. However, a brief examination of the findings is warranted. Among all classes, channelized flows exhibited the most consistent and high predictive performance (R² = 0.73–0.90).

The channelized-obstructed class, although still associated with favorable predictive performance, revealed a moderate reduction in model accuracy compared to purely channelized flows (R² = 0.47–0.63). In contrast, obstructed flows exhibiting no channelized flow path—displayed significantly reduced explanatory power (R² = 0.39–0.70).

Finally, unobstructed flows, despite their apparent topographic open-slope configuration, presented the lowest model performance (R² = 0.31–0.61). Taken together, these findings suggested that runout path morphology may be a control on the predictability of landslide mobility. These results reinforce the argument, supported by Legros (2002) and McDougall (2017), that integrating information on flow confinement and physical obstructions may improve prediction ability of landslide behavior, particularly in urbanized mountainous landscapes such as Petrópolis [20,29].

5.4. Limitations and Future Directions

The results of cross-validation (Table 5) indicate that even the final model has low generalization performance, as evidenced by the low R² value and high error metrics. This trend likely reflects the problem of overfitting, which is often seen in empirically derived landslide prediction models [1,6,12,81]. As we mentioned in the introduction, we consider that the accuracy of landslide runout distance prediction is influenced by two factors: the modeling method and the input data. Based on this, there are two possible reasons for the low performance: first, several effects cannot be described as linear relationships; second, several variables are missing, even though these variables have a significant impact on landslide travel distances. Previous studies have shown that although morphometric indicators capture trends within datasets, their predictive transferability is constrained by the inherently non-linear and site-specific nature of runout processes, particularly in dense urbanized settings [21,29,30,115]. These findings underscore the limitations of multivariate linear regression for runout prediction in complex environments. To improve model performance, we have to test non-linear statistical methods and/or physical models in the future.

In terms of data collection, this study faced four significant limitations. Firstly, although this study prioritized the use of 30 m spatial resolution data for landslide research, focusing on the availability of data immediately after disasters and the ease with which historical data could be accessed across various regions, it is clear that a 30-m resolution may be too coarse to accurately represent the topography and detailed land use in mountainous areas [104,116]. Further consideration of an appropriate spatial resolution is therefore essential. Secondly, detailed data on landslide volume is lacking. Many previous studies have confirmed that landslide runout distance depends heavily on landslide volume e.g., [64,72,73,79,80]. However, while this study obtained data on average landslide depth, it did not obtain depth data for each individual landslide area. This presents a significant challenge given that collecting detailed landslide depth data is often difficult in many areas [2]. Furthermore, although remote sensing has provided more surface information, data on subsurface conditions remains limited. This may also have affected the accuracy of the predictions. Fourth, it is possible that an insufficient number of data may have had an impact of the model performance. It is considered that further data accumulation is necessary. In conclusion, higher prediction accuracy and stability can be achieved by collecting more data on site conditions, landslide morphology and landslide sediment characteristics, and combining this with non-linear statistical methods and/or physical models.

5.5. Practical Applications

The consistent identification of collapse length as the most robust predictor of runout (Section 5.1) suggests its use as a rapid screening metric for preliminary hazard assessments. model performance improved substantially when geoenvironmental data from multiple landslide stages were integrated, underscoring the need to consider collapse, transport, and deposition conditions jointly (Section 5.2). The results also showed that in highly urbanized and geomorphologically complex contexts, remote sensing and field survey data exerted greater influence on predictive performance. Additionally, stratification by runout path typology (Section 5.3) further revealed that channelized flows are the most predictable, whereas obstructed and open-slope flows showed lower predictability. Taken together, these findings demonstrate the practical potential of the models for supporting disaster risk reduction. Despite remaining challenges (Section 5.4), the results provide actionable insights for disaster risk management in resource-constrained contexts based on the developed data collection methods. First, it is essential to recognize that the impact area of a landslide depends on two main factors: magnitude and topographical conditions. When accurate magnitude estimates are not feasible, municipalities can develop multiple scenarios of varying magnitudes to delineate potential hazard zones. Integrating collapse length with susceptibility maps provides a practical approach to identifying slopes with higher potential reach and prioritizing them for detailed surveys or preventive interventions [11]. Second, in terms of site conditions, it is crucial to collect data not only from the initiation area but also from the flow path and deposition zone. Our results confirmed that incorporating deposition-zone parameters yielded the most stable predictive performance. Moreover, utilizing freely available remote sensing data provides an effective means of gathering site-condition information where field surveys are limited. The systematic accumulation of historical data on landslide shape, extent, and geoenvironmental conditions remains crucial for decision-making. This approach is particularly relevant in Brazil, where federal risk assessment methods remain largely heuristic and national inventories continue to serve as the primary basis for producing local hazard maps [1,3,11,20]. For instance, differentiating runout path highlights confined valleys and gullies as priority locations for infrastructure protection, while obstructed and open-slope flows require complementary field validation before supporting land-use decisions [2,3,16,21]. Third, prediction accuracy is not always guaranteed, even with sufficient data on magnitude and site conditions. It is therefore essential to account for uncertainty explicitly when creating hazard maps. One practical strategy is to delineate impact areas broadly enough to encompass the majority of past events, reducing the likelihood of unforeseen disasters in unexpected areas. Within this context, the proposed workflow offers clear benefits by enabling more detailed, evidence-based, and context-specific assessments, while remaining feasible for municipal civil defense agencies.

6. Conclusions

Predicting landslide runout is particularly challenging in tropical mountainous regions, where urban expansion, complex terrain, and scarce data hinder model development. While satellite remote sensing has advanced susceptibility mapping, its use in runout prediction remains limited, especially in Brazil’s urbanized slopes where physically based models are often unfeasible. This study applied a statistical-empirical approach integrating morphometric descriptors with geoenvironmental variables from remote sensing and field surveys across all landslide stages—collapse, transport, and deposition—to evaluate their predictive value and improve understanding of mobility patterns in complex urban tropical settings.

The results confirm that runout distance is not governed solely by the magnitude of the landslide, but is strongly conditioned by a range of geoenvironmental factors that vary across different landslide zones. A key finding of this study is that landslide runout predictions improve markedly when site-condition variables derived from satellite remote sensing and field surveys are incorporated. Models relying solely on post-failure morphometric parameters showed limited explanatory power (R² = 0.06–0.10), underscoring the inadequacy of single-source predictors to represent the multifactorial nature of landslide mobility. By contrast, models integrating remote sensing and field-based variables—especially when combined with morphometric descriptors such as collapse length (Lc)—achieved significantly higher performance (R² = 0.49–0.51). Furthermore, the inclusion of data from all three process zones (collapse, transport, deposition) substantially enhanced prediction accuracy. While single-stage models underperformed (R² < 0.30), multistage combinations (CT, CD, TD) yielded stronger results (R² ≈ 0.44), with full integration across all stages (CTD) producing the best outcomes (R² = 0.49–0.51). These findings reinforce that landslide runout is governed by collective, stage-dependent mechanisms that require comprehensive representation. Furthermore, stratifying models by runout path morphology revealed that channelized flows enabled the most accurate and stable predictions (R² = 0.73–0.90), due to directional confinement and kinetic efficiency. In contrast, obstructed and unobstructed (open-slope) flows—more susceptible to energy dissipation and topographic disruptions—exhibited lower and more dispersed predictive power (R² = 0.39–0.61), emphasizing the importance of accounting for terrain variability in empirical modeling.

Overall, these findings demonstrated that robust landslide runout prediction requires an integrative strategy combining geoenvironmental data (landslide magnitude, remote sensing–derived data, field survey observations), multistage zoning of landslide processes, and consideration of runout path characteristics. This approach not only improves model performance but also ensures practical applicability in hazard assessment frameworks, particularly in data-scarce settings. Although this study advances empirical runout prediction, future research should explore non-linear modeling approaches to more effectively capture landslide behavior, incorporating dynamic variables such as rainfall thresholds and land cover change to strengthen early warning systems and territorial planning in landslide-prone regions.