Volume 126, Issue 5 e2020JF006054
Research Article
Free Access

Sediment Mobilization by Hurricane-Driven Shallow Landsliding in a Wet Subtropical Watershed

C. E. Ramos-Scharrón

Corresponding Author

C. E. Ramos-Scharrón

Department of Geography & the Environment, The University of Texas at Austin, Austin, TX, USA

Correspondence to:

C. E. Ramos-Scharrón,

[email protected]

Search for more papers by this author
E. Y. Arima

E. Y. Arima

Department of Geography & the Environment, The University of Texas at Austin, Austin, TX, USA

Search for more papers by this author
A. Guidry

A. Guidry

Department of Geography & the Environment, The University of Texas at Austin, Austin, TX, USA

Search for more papers by this author
D. Ruffe

D. Ruffe

Department of Geography & the Environment, The University of Texas at Austin, Austin, TX, USA

Search for more papers by this author
B. Vest

B. Vest

Department of Geography & the Environment, The University of Texas at Austin, Austin, TX, USA

Search for more papers by this author
First published: 09 April 2021
Citations: 6


The role of individual tropical cyclones in mobilizing sediment by shallow landsliding has been studied widely in islands of the Pacific but hardly within the Insular Caribbean. An opportunity to conduct such a study materialized in 2017 when Hurricane María provoked over 70,000 landslides on the island of Puerto Rico. Through aerial photo interpretation and high-resolution digital elevation models, this study provides an estimate of the net mobilization and delivery of sediment by shallow landslides in a wet subtropical and actively cultivated 43.8 km2 watershed. Landslide mobilization and delivery are contextualized here in terms of other active sources of sediment and annual sediment delivery. Maximum hourly and 24-h average rain intensities within the study watershed during Hurricane María were 68 and 10.2 mm h−1, respectively, and these represent intensities with local recurrence intervals of less than 10 years. However, these rain intensities were sufficient to trigger 2,318 landslides that mobilized 230,510–436,330 Megagrams of sediment, an amount that is much greater than surface erosion contributions and represents the equivalent of 0.5–3.6 years of annual watershed-scale sediment delivery. The high susceptibility to landsliding in this region of Puerto Rico is linked to the abundance of roads (∼21 km km−2), particularly those crisscrossing terrain with slopes ranging from 30° to 60°. Only 15% of these roads are within actively cultivated areas. The remaining roads represent the support infrastructure of abandoned coffee farms that still appear to be inducing a legacy of much geomorphological and watershed management relevance.

Key Points

  • Landslide sediment mobilization due to a Category 4 hurricane was significant relative to surface erosion and watershed sediment delivery

  • About 77% of the total sediment mobilized by landslides originated within abandoned coffee farms presently under a secondary forest cover

  • A dense road network supporting past and present agricultural activity makes this landscape very susceptible to landslides

Plain Language Summary

Managing sediments is vital for many steep tropical islands where the protection of vital natural resources such as soil, water, and coral reefs is of utmost importance. Wherever rain-driven shallow landslides are prevalent, they tend to be the dominant process releasing sediment from hillslopes and delivering them to streams. This study uses information captured as a response to Hurricane María in 2017, a tropical cyclone that had catastrophic consequences for several Caribbean islands. Aerial photographs and high-resolution topographic data that became available for Puerto Rico after the hurricane allowed us to estimate the volume of sediment released by landslides and evaluate the factors contributing to their occurrence. Results indicate that the coffee-growing region of Puerto Rico is very susceptible to landslides and that these landslides represent a large portion of the sediment impacting its natural resources. The high level of vulnerability to landslides is largely due to the presence of a dense road network that partly serves active farming but that is mostly within presently abandoned farms. The findings of this study imply that management strategies for this region must not only attend active cropland but also abandoned farms to effectively reduce the adverse impacts of erosion.

1 Introduction

1.1 Sediment Releases Associated With Tropical Cyclones

Watershed-scale sediment delivery is the combined result of sediment release from hillslopes, hillslope-to-stream connectivity, and the transport and storage capacity of the fluvial system (Brierley et al., 2006; Church, 2017; Walling, 1983). Sediment delivery is strongly dependent on climatic controls on rock weathering, vegetation cover, and rainfall patterns (Jansen & Painter, 1974; Langbein & Schumm, 1958). Topographic relief and form are relevant in that they control the erosive energy of overland flow (Julien & Simons, 1985), mass wasting potential (Mondini et al., 2013), sediment connectivity (Baartman et al., 2013), and fluvial transport (Mulder & Syvitski, 1996). Lithology also is relevant with the general understanding that young sedimentary deposits are more erosion prone than older crystalline rocks (Syvitski & Milliman, 2007). Land use may alter sediment yields by accelerating surface erosion (Walling, 1999), enhancing connectivity (Lexartza-Artza & Wainwright, 2011), inducing stream channel geometric readjustments (Liébault et al., 2005), and increasing landslide incidence (Glade, 2003). The location and trapping efficiency of man-made dams is another factor that has affected the rate of sediment delivery to the world’s oceans (Syvitski et al., 2005).

Factors leading to high sediment yields tend to coalesce on high-standing tropical and subtropical islands (Milliman et al., 1999). The combination of steep, geologically young, and tectonically active terrain, limited sediment storage opportunities, and oftentimes extensive land use lead to sediment yields commonly in the upper 100 and 1,000 s Mg km−2 yr−1 range, and even up to the 10,000 s Mg km−2 yr−1 (Douglas & Guyot, 2005; Kao & Milliman, 2008; Milliman, 1995; Milliman & Syvitski, 1992). Watersheds with the highest annual sediment yields are located on islands within tropical cyclone corridors (Milliman & Farnsworth, 2013). Despite their limited surface area, high-standing islands represent a significant contribution to the delivery of sediment (Milliman & Meade, 1983) and organic carbon (Bao et al., 2015; Lyons et al., 2002) to the Earth’s oceans.

Some of the largest magnitude sediment transport events on Earth have occurred on high-standing islands during tropical cyclones (Galewsky et al., 2006; Korup, 2012). The geomorphic efficiency of these events is due to the associated extreme rainfall intensities (∼80–150 mm in 1 h; Geerts et al., 2000; Lee et al., 2006; Saito et al., 2014) and daily totals (∼200–1,095 mm d−1; Chien & Kuo, 2011; Kim et al., 2006; Nugent & Ríos-Berríos, 2018; Terry & Raj, 1999). Their geomorphic effectiveness is not only associated with high fluvial transport capacity (Kostaschuk et al., 2003), but also to the activation of stochastic processes such as gullying, stream channel scouring, and shallow landsliding (Gupta, 1999; Page, Reid, & Lynn, 1999). Some notable shallow landslide-related sediment mobilization values associated with individual tropical cyclones include the ∼4,700 Mg km−2 recorded in New Zealand in the Tutira watershed during Cyclone Bola in 1988 (Page, Trustrum, & Dymond, 1994), and the ∼130,500 and ∼233,000 Mg km−2 reported for Taiwan in the Tsengwen Reservoir and Kaoping River watersheds (respectively) during Typhoon Morakot in 2009 (Y.-C. Chen et al., 2013). Assuming an annual sediment yield for the Tutira watershed of 6,570 Mg km−2 yr−1 (Hicks et al., 2011), sediments released by landsliding during Cyclone Bola represent about 0.7 years of sediment delivery. In contrast, sediment releases by landsliding for the Taiwan case match 20–25 years of sediment yields (annual yields of 5,975 and 9,600 Mg km−2 yr−1 for the Tsengwen and Kaoping catchments, respectively; Y.-C. Chen et al., 2013).

The Insular Caribbean has been underrepresented in worldwide sediment yield assessments. Only two watersheds from the entire region figure in the Global River Sediment Yields Database (Food and Agriculture Organization, 2016). This is in part because most Caribbean islands have never developed a long-term sediment yield monitoring program (UNEP, 1994). Sediment yield estimates in the region have relied on indirect measures such as isotopic signals of sediment samples (e.g., Brown et al., 1998; Nagle, Lassoie, et al., 2000; Rad, Rivé, et al., 2013), rarely validated watershed modeling efforts (e.g., Febles-González et al., 2012; Korman et al., 2020), sedimentation rates from terrestrial or marine sinks (e.g., Bégin et al., 2014; Nagle, 2001), or short-lived suspended sediment monitoring efforts (e.g., Ramos-Scharrón & MacDonald, 2007). Some documented rates for the region include those for the Dominican Republic (1,960–4,575 Mg km−2 yr−1; Nagle, 2001), several from the Lesser Antilles (800–4,000 Mg km−2 yr−1 for Guadalupe, Martinique, and Dominica; Rad, Louvat, et al., 2006) and Puerto Rico (PR) (120–4,300 Mg km−2 yr−1; Larsen & Webb, 2009). Quantification of sediment yields is of utmost concern for the islands because of water reservoir sedimentation (Gellis et al., 2006; Nagle, Fahey, & Lassoie, 1999) and the potentially deleterious effects on coral reef ecosystems (Orlando & Yee, 2017). Anecdotally, surface erosion from agricultural lands has been considered a key source of sediment (e.g., Rawlins et al., 1998). However, surface erosion only accounts for ∼6%–18% of sediment yields (Nagle, 2001; Ramos-Scharrón & Figueroa-Sánchez, 2017). Quantifying the contribution from shallow landslides to sediment yields is therefore essential to better understand its impact on total sediment yield (Ahmad et al., 1993), particularly given the quantifiable effect of climate change on tropical cyclone rainfall (Lim et al., 2018; Murakami et al., 2018) and its potential for greater landslide incidence (Chiang & Chang, 2011).

Shallow landslides are common to the steep portions of Caribbean islands, particularly during tropical cyclones (DeGraff et al., 1989). PR is no exception as tens to hundreds of landslides have been documented following tropical cyclones in 1985 (Tropical Storm Isabel; Jibson, 1989) and 1989 (Hurricane Hugo; Larsen & Torres-Sánchez, 1992; Scatena & Larsen, 1991). One-hour and daily rainfall thresholds for shallow landslide initiation in PR are in the 50–90 mm h−1 and 110–160 mm d−1 range, respectively (Larsen & Simon, 1993; Pando et al., 2005). Prominent landslide occurrence on the island has been associated with slope gradients ranging from 30° to 45° (Larsen & Parks, 1998; M. C. Larsen et al., 2004), lithologies with volcanic, volcaniclastic, hydrothermally altered, and intrusive origins (Monroe, 1979), and specific land cover classes such as cropland, pasture, and built-up areas (Larsen & Torres-Sánchez, 1998; Lepore et al., 2012). The influence of roads in slope instability has been crucial in some areas (Larsen & Parks, 1997), but has been ambiguous in others (Lepore et al., 2012). Shallow landslides have been found responsible for 92%–99% of the sediment produced from both forested and actively cultivated areas of eastern PR (Larsen, 2012). However, no study has evaluated the role of landslides in the sediment budget of central-western portions of the island where agriculture, largely in the form of sun-grown coffee cultivation, remains prevalent.

An opportunity to study sediment mobilization by landsliding arose when Hurricane María (HM) entered PR on September 20, 2017 as a borderline Category 4 major hurricane (Pasch et al., 2018; Figure 1). The damage induced by HM on the island was catastrophic (Lugo, 2019), and it was in part related to the intense rainfall registered mostly within an 8-h time span (up to 740 mm; Ramos-Scharrón et al., 2020). Daily rainfall intensities during HM surpassed the 100-year recurrence interval threshold for 34% of the long-term weather stations (Keellings & Hernández-Ayala, 2019) and for about 55% of the entire island’s landmass (Ramos-Scharrón & Arima, 2019). At an island-wide scale, HM ranks first in daily rainfall intensities among 60 of the most important tropical cyclones that have affected PR since 1899 (Ramos-Scharrón & Arima, 2019). HM rainfall triggered a population of more than 70,000 shallow landslides (Hughes et al., 2019) with a spatial distribution controlled by slope, lithology and soil type, soil moisture, land cover, proximity to roads, and HM rainfall (Bessette-Kirton, Crovski-Darriau, et al., 2019; Hughes & Schulz, 2020; Ramos-Scharron et al., 2020). However, the geomorphological significance of HM landslides is yet to be evaluated.

Details are in the caption following the image

Study area maps and photographs of landslides. (a) The Lago Lucchetti Watershed (LLW) displaying the location of Lago Lucchetti, the stream network, and the spatial distribution of individual landslides. Insert shows the location of LLW in relation to Puerto Rico and Hurricane María’s track. (b) Close-up view of portions of LLW displaying the location of landslides categorized by type. Examples of (c) cutslope, (d) fillslope, and (e) non-road-related landslides found within LLW, respectively.

This study represents an assessment of the quantity of sediment released by HM landslides in an actively cultivated watershed in western PR. The specific objectives of this study were to (1) develop a power-law relationship between planimetric landslide scar area and volume; (2) contextualize the total landslide-related sediment mobilized and potentially delivered to streams during HM in relation to surface erosion and watershed-scale sediment delivery; and (3) characterize the importance of factors controlling landslide-related sediment releases during HM.

1.2 Study Area

The municipality of Yauco in western PR has been one of the most prominent coffee producing areas of PR since Spanish colonial times (Baralt, 1984; O’Neill, 1990). With elevations ranging from 150 to 970 m, mean annual temperatures of 21°C–24°C, and an annual rainfall of ∼ 1,320–2,200 mm y−1 (Daly et al., 2003), the 43.8 km2 Lago Lucchetti Watershed (LLW) provides optimal climatic conditions for coffee cultivation, particularly at its highest elevations. LLW also is part of the island’s Southwest Water Project, a series of interconnected reservoirs built to generate hydroelectric power and to convey water to the drier southwest region of PR (Ortiz-Zayas & Terrasa-Soler., 2001). In recent years, LLW has gained prominence as a priority water and soil conservation area due to the impacts of its sediments on the coral reefs near Guánica Bay (Smith et al., 2017). Throughout the early 21st century, about 8%–30% of the watershed has been agriculturally active, most of it devoted to coffee in the sun-grown cultivation modality (Gould et al., 2008; NOAA, 2015). LLW is within PR’s wet and moist forest life zone (Ewel & Whitmore, 1973) and is underlain mostly by volcaniclastic sandstones and siltstones of the Yauco formation (McIntyre, 1975) and Ultisols of the Maricao and Agüeybaná series (Gierbolini, 1975).

The combined effect of largely exposed and highly erodible soils typical of sun-grown coffee cultivation, high annual rainfall, and steep topography are presumed to be the reason why sediment yield rates into Lago Lucchetti (2,800–4,200 Mg km−2 y−1; Gómez-Fragoso, 2016; Soler-López, 2001a) are the highest amongst all 13 major water reservoirs on the island (PR-wide range: 420–2,600 Mg km−2 y−1; Soler-López, 2001b). Curtailing surface erosion has been the focus of watershed management strategies for LLW (Sturm et al., 2012). However, surface erosion from active farms can account for only 6%–18% of average sediment yields in LLW (Ramos-Scharrón & Figueroa-Sánchez, 2017). The amount of sediment imported to LLW from three reservoirs in the Río Grande de Añasco watershed through the system of interconnected tunnels is unknown (Yuan et al., 2016). Similarly, unknown is the contribution from other sources of sediment such as landslides.

Landslide vulnerability maps indicate that ∼71%–78% of LLW has a high risk of landsliding (Hughes & Schulz, 2020; Lepore et al., 2012), and this proved to be true during HM. Initial mapping efforts identified 1,213 landslides within LLW (Hughes et al., 2019). The maximum landslide density in LLW was 176 slides per km2 and this in the moderate-to-high range with respect to the rest of the island (island-wide mean and maximum were 11.2 and 432 slides per km2, respectively; Ramos-Scharrón et al., 2020). HM rain total within LLW was 125–250 mm with maximum 24-h average rain intensities of 5.1–10.2 mm h−1 (Ramos-Scharrón & Arima, 2019). Rain intensities near LLW reached hourly intensities of 68 mm h−1 two times, once after a cumulative rain total of 103 mm had already been registered and another after 289 mm (Río Yauco-USGS station, only 5 km south of Lago Lucchetti; Figure 2). However, neither hourly nor 24-h intensities were out of the ordinary for this region. Maximum 1-h rain intensities in LLW were within the range of 1-year recurrence intervals (54–81 mm h−1); recurrence intervals for 24-h intensities ranged from 1 to 10 years (4–11 mm h−1; Stations: Yauco 1NNE, Yauco 1NW; Maricao, Maricao 2SSW, and Maricao Fish Hatchery; Bonnin et al., 2006).

Details are in the caption following the image

Hourly rain intensities registered at Rio Yauco station (USGS 50126150) for September 19–21, 2017 (green solid line). Dashed orange line represents the 1-h rain threshold for shallow landslide initiation from Pando et al. (2005).

2 Materials and Methods

2.1 Landslide Database Generation and Lidar-Based Analyses

A new polygon geodatabase of shallow landslide scars was generated for LLW through on-screen digitizing. Polygons were delineated based on georeferenced images taken soon after HM (FEMA; Quantum Spatial, Inc.; September 25–27, 2017; ∼15-cm pixels). It is very unlikely that any of the fresh landslide scars observable in these images were triggered by Hurricane Irma just 2 weeks before HM (September 5–7, 2017), as no landslides were observed during a field visit to the area between the storms. Additionally, LLW received only 70–120 mm of total rainfall with 24-h maximum intensities from 43 to 79 mm d−1 during Hurricane Irma (Ramos-Scharrón & Arima, 2019). These intensities did not exceed the daily rainfall intensity threshold for the island (110–160 mm d−1; Larsen & Simon, 1993; Pando et al., 2005). Furthermore, maximum 1-h intensities at the Río Yauco station during Hurricane Irma reached only 4.6 mm h−1.

Digitized polygons exclusively delineated landslide source areas from their head scarps to a visible lip noted by sharp differences in color that represent clear transitions from bare to vegetated surfaces. Mapped landslides included roadcut and fillslope failures, landslides in the vicinity of streams, and those both within actively used and forested areas. Landslide scars were categorized into those involving road cutslopes or fillslopes and those that appear unrelated to roads (labeled as “non-road”). Cutslope failures were mostly of the debris slide type, while fillslope and non-road scars were mostly hillslope debris flows or avalanches (Hungr et al., 2014). For landslides that triggered a debris flow, the scour feature on the hillslope was also digitized. Scar polygons did not include deposits. Given the high resolution of the imagery and the visibility allowed by widespread defoliation, the smallest landslide scar digitized had a surface area of 4 m2. Initial landslide mapping efforts depended solely on the aerial images, but subsequent work also relied on an existing landslide point database (Hughes et al., 2019) and a pair of lidar-based digital elevation models (DEMs) representing both pre-HM (2016–2017) and post-HM (2018) conditions. The US Geological Survey 2016–2017, 1-m gridded, hydro-flattened DEM was derived with two airborne lidar systems, Riegl 680i and Riegl 780, flown in two campaigns (January 26 to May 15, 2016 and December 8, 2016 to March 16, 2017). This project used nominal pulse spacing of 0.7-m (i.e., 1 lidar point per 0.7 m) and had horizontal and nonvegetated vertical accuracies of ±21.6 and ±18.5 cm, respectively. The 2018 1-m gridded hydro-flattened DEM collected data with two Riegl VQ1560i lidar sensors. The lidar was collected in one campaign in late 2018 and used a nominal pulse spacing of 0.35-m and had horizontal and nonvegetated vertical accuracies of ±34.1 and ±13.4 cm, respectively. Both DEM data sets were produced to meet the ASPRS Positional Accuracy Standards for Digital Geospatial Data (2014).

The two DEMs were used to calculate a 1-m resolution elevation difference raster calculated between the 2018 and 2016–2017 DEMs (i.e., 2018 minus 2016–2017) Therefore, elevation losses resulting from landsliding show as negative values in the resulting raster. A similar approach was used to estimate landslide volumes in Taiwan (S.-C. Chen et al., 2019; Tseng et al., 2013). Many polygon boundaries mapped from aerial image interpretation were subsequently edited to improve their fit in relation to the difference raster and this involved both enlarging and shrinking the size of scar polygons. Only scars that displayed distinct visual cues in the aerial imagery of obvious landsliding that also displayed a distinctive elevation loss according to the difference raster were included in our inventory. The difference raster could not reveal elevation losses for all landslide scars because either landslide erosion may have been only very superficial and thus undetectable or because of evident inaccuracies in portions of the 2018 DEM. These problem areas were constrained to specific and mostly forested hillslopes that displayed the most severe level of wind damage in the post-HM aerial imagery. Therefore, we presume that the downed trees might have prevented an accurate calculation of the land surface with the ground filter algorithm applied to the lidar signals (Meng et al., 2010). Therefore, the polygon geodatabase had an attribute to identify scars for which we could rely on the elevation difference raster to calculate volume. Landslides from ∼80% of all scar polygons displayed what were considered reliable elevation losses and were therefore used to calculate landslide volume as:

where V is volume in m3, A is planimetric area of each landslide scar in m2, and urn:x-wiley:21699003:media:jgrf21346:jgrf21346-math-0002 is in meters and stands for the absolute value of the average loss in elevation between the 2018 and 2016–2017 rasters for each landslide scar polygon. The mean depth for each scar was calculated using the Zonal Statistics as Table tool in ArcGIS 10.7.1. We calculated not only a single volumetric value based on the urn:x-wiley:21699003:media:jgrf21346:jgrf21346-math-0003 but also a range of volume values for each scar to account for the error in the vertical accuracy of the two lidar elevation models (urn:x-wiley:21699003:media:jgrf21346:jgrf21346-math-0004).

The resulting scar volumes were used to develop the power-law relationship between area and volume as follows:
where V is simply the volume based on the urn:x-wiley:21699003:media:jgrf21346:jgrf21346-math-0006 value captured from the Zonal Statistics as Table tool. Equation 2a is as follows in its Log10 transformed form:

where α and γ are fitting parameters and V and A are the same as for Equation 1. Regression parameters were fitted in MATLAB following a robust-linear fitting method as recommended by Guzzetti et al. (2009) to minimize the effects of outliers. Volumes for the remaining 20% of landslide scars for which no direct volume estimate was possible were estimated based on regression results which accounted for the 95% confidence interval (CI) estimate of resulting parameter values. The similarity of scar area, depth, and volumes for the three types of landslides mapped (i.e., cutslope, fillslope, and non-road) was tested based on the nonparametric Kruskal-Wallis ANOVA test and Dunn’s posthoc test due to the heteroscedastic nature of the values (Zar, 2010). The summed volume and mass of sediment is referred to here as the amount mobilized (i.e., released) by landslides (a la Guzzetti et al., 2009).

2.2 Sediment Budget Analyses

A sediment delivery ratio approach was used to approximate the proportion of mobilized sediment that had the potential of being delivered to the stream network as a direct consequence of the mass movements themselves. This procedure provides a rough estimate on the degree of coupling (Persichillo et al., 2018) or connectivity (Li et al., 2016) between landslide sources and the stream network. The approach is based on empirical approximations of landslide runout lengths that are solely based on landslide volume (Legros, 2002). Two different equations were used. One was originally developed for landslides that did not transform into debris flows (Hürlimann et al., 2015) as:
where L is maximum runout length in meters and V is landslide volume in m3. This equation delimited relatively short runouts and was applied exclusively to cutslopes in LLW which are typically of the debris slide type (Hungr et al., 2014). A second equation was derived from debris flows (Bathurst et al., 1997) and was applied here to both fillslopes and non-road landslide types in LLW:
The estimated amount of sediment delivered to the stream network from each individual scar is a function of runout distance (L) and the downslope distance (D) to the stream network following Bathurst et al. (1997) and Cislaghi and Bischetti (2019):

where SD is the estimated sediment delivery to streams in m3. Scar polygon centroids were used as point sources for the Cost Path as Polyline tool in ArcGIS 10.7.1 to determine downslope distance to streams (the 1-m DEM and flow direction rasters served as the cost distance and cost backlink inputs, respectively). Path lengths were delimited by a 25-m buffered version of the stream network using the Erase tool. The stream database was developed using the 1-m DEM raster using 5 ha as a source area stream threshold value, which generated a network that is consistent with our field observations on the extension of stream channels in LLW. Equation 5 presumes that deposition of debris flow material over the runout distance is uniform, following Cenderelli and Kite (1998). However, the approach is likely to underestimate delivery because it does not consider the importance of landslide coalescence in extending runout distances, which was identified as an important factor for HM-triggered landslides in PR (Bessette-Kirton, Coe, et al., 2020).

The amount of sediment mobilized and delivered by HM-triggered landslides was compared to annualized surface erosion rates from both forested areas and coffee farms. Annual sediment production rates from these areas assumed that 8%–30% of LLW is covered by coffee farms and that the remaining surface area is in forests. Sediment production rates for coffee farms ranges from 1,200 to 1,800 Mg km−2 y−1 (Ramos-Scharrón & Figueroa-Sánchez, 2017; Ramos-Scharrón & Thomaz, 2017), while those from forested surfaces ranged from 20 to 50 Mg km−2 y−1 (Figueroa-Sánchez, 2019). Due to the fact that landslide mobilization and delivery estimates derived here are for a single event, no annualized comparisons are possible. Therefore, comparisons were based on the equivalent number of years the mobilized and delivered sediment represents by dividing the mass of landslide sediment by the watershed scale surface erosion rates and LLW-scale annual sediment delivery rates (Gómez-Fragoso, 2016; Soler-López, 2001a).

2.3 Landslide Controlling Factors: Geodatabases and Frequency Ratio Analyses

Factors known to have determined the location of HM landslides throughout PR include event total rainfall, slope, lithology, soils, land cover, and roads, amongst others (Bessette-Kirton, Crovski-Darriau, et al., 2019; Hughes & Schulz, 2020; Ramos-Scharrón et al., 2020). The effects of these factors in the spatial distribution and areal extent of landsliding in LLW were empirically evaluated in this study through GIS analyses. A brief description of the geodatabases used for these analyses is shown in Table 1 and supporting information. These included an interpolated HM rainfall total raster (Ramos-Scharrón & Arima, 2019), and the 1-m resolution DEM from 2016 to 2017 for determining elevation, slope, and the extent of the stream network. Additional databases included publicly available geologic terrane, soil series, and land cover layers. The roads layer was a combination of a publicly available version of an official state road map with additional roads added by on-screen digitizing of a hillshade version of the 2018 1-m resolution lidar DEM. The lidar data allowed for the visual identification of roads obscured by dense forest canopies in aerial imagery as done elsewhere (White et al., 2010).

Table 1. Tabulated Description of the Spatial Databases Used for Analyzing the Importance of Controlling Factors in Mobilizing Sediment by Landsliding
Geodatabase Source(s) GIS analyses General attribute description
HM total rainfall 90-m raster (Ramos-Scharrón & Arima, 2019) Resampled to 1-m resolution; zonal statistics (mean) 123–249 mm [178 mm; mean]
Elevation Gridded 1-m 2016–2017 DEM derived from USGS product (link) Zonal statistics (mean) 147–965 m [500 m]
Slope Gridded 1-m 2016–2017 DEM derived from USGS product Slope (degrees); zonal statistics (mean) 0°–80° [32°]
Geologic terrane Rasterized version of Bawiec (1998) Resampled to 1-m resolution; zonal statistics (mean) Intrusive (27.3%), Sedimentary (71.6%), and Metamorphic (1.1%)
Soils Vectorized version of PR’s soil series map (https://websoilsurvey.nrcs.usda.gov/) Rasterized to 1-m resolution; zonal statistics (majority) Quebrada clay loam (6.3%), Caguabo clay loam (58.9%), Agüeybaná clay loam (11.2%), Humatas clay (23.1%), Montegrande clay (0.1%), and Malaya clay (0.3%)
2010 Land Cover 30-m resolution land cover map (https://coast.noaa.gov/digitalcoast/data/) Resampled to 1-m; zonal statistics (majority) Built up (2.2%), cropland (8.1%), grassland (3.9%), forest (78.0%), and shrub (7.8%)
Roads TIGER/Line vectorized 2015 primary and secondary roads (www.data.gov); on-screen digitized lines based on hillshaded version of 1-m DEM for all other roads Minimum Euclidian distance (scars to roads) 0–306 m (26 m)
Streams Stream network derived from 1-m USGS DEM Flow direction; flow accum.; 5 ha threshold; Zonal statistics (mean); Euclidean distance 27–377 m (153 m)
  • Note. Also listed are the GIS analyses conducted to assign values from each of the databases to individual landslide scars. See suppporting information for additional information on these databases.
  • Abbreviation: DEM, digital elevation model.
Spatial associations of potential controlling factors and landslide incidence were conducted using frequency ratios (FRs). FR, also known as the ratio of probability (Marc et al., 2018), has been widely used in landslide susceptibility studies because of its simplicity, ease of implementation, and overall good performance in explaining factors associated with the spatial distribution of landslides (e.g., Lepore et al., 2012). In an ordinary FR, factors are evaluated individually by establishing classes for each of the factors (i.e., bins) and applying this formula:

where FRim is the frequency ratio for factor i in bin m, Lim is the area of landslide scars in bin m of factor i, LT is the total area of all landslides, Aim is the total area of factor i in bin m within the entire study area, and AT is the total area for LLW. The denominator, urn:x-wiley:21699003:media:jgrf21346:jgrf21346-math-0012, is referred to as urn:x-wiley:21699003:media:jgrf21346:jgrf21346-math-0013 or the probability of factor i in bin m over the entirety of the study area. The 95% CI for every FR value was calculated by providing a range of urn:x-wiley:21699003:media:jgrf21346:jgrf21346-math-0014 values as described by Marc et al. (2018): urn:x-wiley:21699003:media:jgrf21346:jgrf21346-math-0015 where N equals the total number of landslide scars in LLW. When conditions yield FR values that are significantly greater than 1, this implies that there is an overabundance of landslide scar area within the bin m of factor i than the relative abundance of that factor in the entire watershed. In other words, FR values exceeding 1 represent characteristics that showed a high propensity to landsliding. The opposite is true for characteristics with FR values significantly less than 1.

We also relied on a joint frequency ratio (JFR) between two factors to simultaneously evaluate landslide propensity to the combined effects of land cover, slope, and distance from roads. In JFR, factors i and j are defined similarly as for FR, except that in this case, the classes are unique combinations of bins of two factors simultaneously. For instance, the JFR of factor slope with m bins with factor distance to roads classified into n bins will produce m × n joint bins, assuming we observe events in all bins. For example, if we reclassify slope raster cells into bins m = 0°–5°, 5°–10°, etc., and reclassify distance to roads cells into bins n = 0–10 m, 10–20 m, etc., the JFR between slope and distance for the first bin (i.e., slope between 0° and 5° and distance to road is 0–10 m) is:

where urn:x-wiley:21699003:media:jgrf21346:jgrf21346-math-0017 is the area of landslides with slope between 0° and 5° and distance to road within 0–10 m, and urn:x-wiley:21699003:media:jgrf21346:jgrf21346-math-0018 is the net area within LLW having the same characteristics. All other terms are defined as for Equation 6. Therefore, JFRs can aid in the understanding of combined factors that might be associated with higher landslide scar areas and thus higher sediment mobilization. JFRs can be extended to accommodate more than two factors but the number of combinatorial bins will grow exponentially and complicate interpretation. In addition, visual displays are limited to three dimensions or two factors plus the JFR value. In this study, JFRs were calculated from 1-m resolution raster files, yielding over 44 million pixels for the entire LLW area.

3 Results and Discussion

3.1 Landslide Inventory and Area-Volume Power Law

We identified 2,318 fresh landslide scars in LLW using the post-HM imagery. About 26% of the landslides were cutslope debris slides, while 43% and 31% were fillslope and non-road-related scars, respectively. Scar planimetric areas spanned four orders of magnitude (100–103 m2) with a mean of 163 m2. Cutslope scars had the smallest average scar area among the three mapped types (Table 2; Figure 2). A total of 1,852 scar polygons were used for the direct volumetric estimates using the pre- to post-HM elevation difference raster. The average depth of landslide scars was 0.5 m, with individual values ranging from ∼0.0 to 3.9 m. There were no statistical differences in average depths among the three landslide types. Scar volumes ranged over five orders of magnitude (10−1 to 103 m3) with an average of 123 m3. Non-road landslide scar volumes were the largest averaging 156 m3, followed by those for fillslopes and cutslopes (119 and 90 m3, respectively) (Figure 3).

Table 2. Tabulated Description of Landslide Scar Metrics
Landslide type Scar area (m2) Scar depth (m) Scar volume (m3)
Count Min Max Mean Group Count Min Max Mean Group Count Min Max Mean Group
Cutslope 603 4.6 1,841 94 A 482 0.03 3.75 0.66 A 482 0.2 5,697 90 A
Fillslope 1,004 5.4 3,095 167 B 793 0.02 3.92 0.61 A 793 0.1 3,273 119 B
Non-road 711 4.2 3,326 215 C 577 0.02 3.37 0.62 A 577 0.5 8,399 156 C
All 2,318 4.2 3,326 163 1,852 0.02 3.92 0.63 1,852 0.1 8,399 123
  • Note. Groupings are based on nonparametric Kruskal-Wallis ANOVA test and Dunn’s posthoc testing.
Details are in the caption following the image

Cumulative distribution curves of scar dimensions in terms of relative frequency for: (a) scar area [A], (b) average depth [h], and (c) volume [V], and (d) for scar volume relative to total volume.

There were only minor differences in the scaling parameters for the three types of landslides considered in this study. The 95% CIs of the scaling parameters overlapped and therefore were deemed not statistically significant (Table 3; Figures 4a–4c). Consequently, regression analyses based on the entire scar area-volume data set yielded the following equation (Figure 4d):
Table 3. Summary of Scar Area-Volume Regression Results
Landslide type Power law regression results
R2 p-value α 95% CI γ 95% CI
Cutslope 0.637 <0.0001 −0.684 −0.833 −0.534 1.209 1.127 1.291
Fillslope 0.647 <0.0001 −0.692 −0.817 −0.568 1.183 1.122 1.244
Non-road 0.725 <0.0001 −0.513 −0.631 −0.394 1.105 1.049 1.161
All 0.676 <0.0001 −0.610 −0.683 −0.537 1.152 1.116 1.188
Details are in the caption following the image

Power-law landslide area-volume relationships for (a) cutslope, (b) fillslope, (c) non-road-related, and (d) all landslide scars within LLW. LLW, Lago Lucchetti Watershed.

This equation converts to:

The 95% CI estimates for α ranged from 0.207 to 0.290, and 1.116 to 1.188 for γ. These values are well within the scaling values reported in the literature (α: 0.074–12.27; and γ: 0.88–1.450; Guzzetti et al., 2009). Using area-volume data from debris flows, soil slips, and slumps from central and eastern PR (Larsen & Torres-Sánchez, 1998), Guzzetti et al. (2009) determined an α = 1.826 and a γ = 0.898 suggesting a higher volume for landslides in the 10−1–102 m2 range than those reported here, yet similar area-volume relationships for scars in the 103 m2 range. The γ = 1.15 for this study agrees with the 1.1–1.3 considered representative of shallow landslides involving mostly soil substrates (I. J. Larsen et al., 2010).

3.2 Sediment Mobilization and Delivery

Equation 8b was applied to the 466 of scar polygons for which the elevation difference raster was deemed inappropriate to determine a volume. The estimated total sediment mobilized by HM within LLW was 272,200 m3 (192,090–363,610 m3 when accounting for the combined raster error in determining urn:x-wiley:21699003:media:jgrf21346:jgrf21346-math-0021 and the 95% CI of Equation 8b). About 17% of the sediment was generated by cutslope failures, 43% from fillslopes, and 40% from non-road-related landslides. To estimate mass, we assumed a dry bulk density of 1.2 Mg m−3 which is typical for Ultisols with an organic content of ∼10% (Heuscher et al., 2005). Our calculation yields 326,640 Mg of sediment (230,510–463,330 Mg), which over a 43.8 km2 yields an area-normalized mobilization value of 7,460 Mg km−2 (5,260–9,960 Mg km−2) or a vertical denudation equivalent of 6.2 mm (4.4–8.3 mm). This is within the range of most per event denudation values reported for Taiwan (0.54–12.9 mm event−1) with the exception of that related to Typhoon Morakot (186 mm event−1; Y.-C. Chen et al., 2015). At LLW, the largest two landslides (5,700 and 8,400 m3) were responsible for only 5% of the total mobilized volume. This contrasts with that reported by Guzzetti et al. (2009) in central Italy, where a few large landslides were responsible for about half of the mobilized sediment.

The median runout distance (L) estimates for cutslopes based on Equation 3 was 9.8 m with values from 1.0 to 126 m. Median distances for fillslope and non-road landslides based on Equation 4 were 78 and 83 m, respectively (range of 14–283 m for fillsopes and 20–376 m for non-road). Median runout lengths for a subsample of HM landslides mapped within volcaniclastic geologic terranes similar to that in LLW were 17–30 m for individual landslides and 46–60 m for landslides that coalesced into single runout paths (Bessette-Kirton, Coe, et al., 2020). It is important to note that these mapped runout lengths are for a combination of debris slides, avalanches, and flows, and that some of the values do not represent maximum values because these were truncated by streams. Therefore, our estimated runout lengths for fillslopes and non-road scars in LLW are in general agreement with those mapped elsewhere in PR.

About 40% of all landslide scars had some potential of delivering sediments according to Equations 3-5 (i.e., L > D). Higher delivery potential was obviously associated with the larger scars due to the dependency of runout lengths on scar volume (Figure 5). Greater delivery was also estimated for fillslope and non-road scars as these were not only larger but also in closer proximity to streams than cutslope landslides. Overall, our approach suggests that ∼119,000 m3 (82,700–163,100 m3) or 43% of all mobilized sediment was estimated to be delivered to the stream network during HM as a direct consequence of the slope failures themselves. About 44% and 55% of this sediment is associated with fillslope and non-road scars (respectively); only 17% is from cutslopes. The 292 landslides that occurred within 25 m of streams were estimated to deliver about 17% of all delivered sediment.

Details are in the caption following the image

Landslide sediment delivery potential estimates. Relationship between landslide scar volume with estimated runout distance (L) (solid black lines), and estimated downslope distance from scar centroid to the 25-m stream buffer (D; points) for (a) cutslope [L based on Equation 3], (b) fillslope, and (c) non-road [Non-rd] scars [L for b and c is based on Equation 4]. Scars plotting above the line represent those for which no sediment delivery is possible because their downslope distance is longer than the estimated runout distance (D > L). (d and e) Landslide scars expected to deliver and not deliver (solid colors and grays, respectively) sediment to streams by (d) number of scars and (e) volume. (f) Net volume of sediment delivered by scar type and volume.

For context, we compare HM landslide mobilization to other sources of sediment active in LLW and to watershed scale sediment delivery rates. Surface erosion from coffee farms within LLW mobilizes from 4,210 to 23,650 Mg y−1 (Ramos-Scharrón & Figueroa-Sánchez, 2017). These surface erosion estimates only represent from 3.4% to 13% of annual sediment yields. Meanwhile, surface erosion from forested portions of LLW is expected to produce an equivalent of only 0.5%–1.1% of watershed-scale sediment output (610–2,000 Mg y−1). Assuming a dry bulk density of 1.2 Mg m−3, HM landslide sediment delivery to streams and total mobilized sediment represents 99,270–195,670 Mg and 230,510–436,330 Mg, respectively. These totals amount to 4–100 and 50–715 years of annual erosion from coffee farms and forested hillslopes, respectively. This emphasizes the vital role that landslides play in the sediment budget of LLW as has been shown elsewhere in PR (Larsen, 2012).

Long-term landslide mobilization rates are typically greater than sediment yields (Broeckx et al., 2020). We believe that this is viable in LLW for the following reasons. First, the mass of sediment delivered and mobilized by landsliding during HM represents from 0.5 to 3.6 years of watershed-scale annual sediment delivery (122,640–183,960 Mg y−1). These values are mostly within the 0.7 and 25 years reported for Cyclone Bola in New Zealand and for Typhoon Morakot in Taiwan, respectively (Y.-C. Chen et al., 2013; Page, Trustrum, & Dymond, 1994), where landslide mobilization rates are indeed greater than watershed-scale sediment delivery. Second, mobilization by landsliding of the magnitude induced by HM (i.e., 230,510–436,330 Mg event−1) would need to occur three to eight times every decade for this sediment to at least match LLW-scale sediment delivery. This implies the need for significant sediment mobilization during relatively frequent storms (i.e., not just extreme events), as has been reported for New Zealand (Reid & Page, 2002) and Japan (Saito et al., 2014). This appears feasible given that HM maximum hourly and daily rain intensities rated as only a 1- to 10-year rainfall event at LLW. Third, assuming that individual landslide scars mobilize ∼100–190 Mg of sediment (mean volumes of all scars accounting for errors were 83–157 m3), landsliding rates of 15–42 slides km−2 y−1 would be required to match LLW yields. In comparison, landslide frequency in pasture-covered areas of New Zealand ranges from 8 to 28 slides km−2 y−1 (Reid & Page, 2002). Finally, long-term landslide sediment production rates for other human-disturbed watersheds in PR range from 4,000 to 5,000 Mg km−2 y−1 (Larsen, 2012). These ranges are well within the moderate to high range of values reported worldwide (0.4–73,200 Mg km−2 y−1; Broeckx et al., 2020) and similar to tropical cyclone influenced New Zealand (530–3,700 Mg km−2 y−1; Reid & Page, 2002), Japan (950–6,200 Mg km−2 y−1; Broeckx et al., 2020; Imaizumi & Sidle, 2007; Koi et al., 2008), and Taiwan (3,200–6,200 Mg km−2 y−1; Y.-C. Chen et al., 2015).

3.3 Factors Controlling Sediment Mobilization by Landslides

The median HM rainfall within LLW was about 170 mm (Ramos-Scharrón & Arima, 2019). However, ∼80% of all landslide mobilized sediment occurred in areas with rain values less than 170 mm (Figures 6a–6c). The median elevation of LLW is ∼480 m, yet 72% of the sediment mobilization occurred at higher elevations (Figures 6d–6f). About 97% of the mobilized sediments occurred on slopes between 30° and 50°, although these slopes correspond only to 61% of LWW (Figures 6g–6i). FR values are consistently above 1 for all slopes greater than 35°, however, slopes steeper than 60° only make up 0.4% of LLW and contribute only 1.4% of the total sediment mobilized. The effect of slope described here is similar to what has been described for the entire population of HM landslides over PR (Ramos-Scharrón et al., 2020). The effect of elevation is likely linked to the effects of slope, land use, and roads as higher elevations are notably steeper and have a higher abundance of roads and both active and abandoned cropland than lower areas. About half of LLW is within 80 m of a stream, but about half of the mobilized sediment was released within 50 m of streams (Figures 6j–6l). However, only areas within 20 m of streams that make up only ∼16.5% of the entire surface area of LLW had FR values that were consistently greater than 1.

Details are in the caption following the image

Evaluation of factors controlling landslide sediment mobilization. Proportion of LLW surface area (%LLW), scar volume, and associated frequency ratios (FRs) in relation to HM rainfall (a–c), elevation (d–f), slope (g–i), Euclidean distance from streams (j–l), and Euclidean distance from roads (m–o). Black points in (c, f, i, l, and o) represent FR values that are statistically distinct from; open circles represent those that are not.

Almost three-fourths of LLW is underlain by volcaniclastic lithologies and although about 60% of the mobilized sediment occurred within this geologic terrane, regions with intrusive rocks proved to have the highest FR value and were thus more landslide prone (Figures 7a–7c). The propensity of intrusive rocks to slide is something that is consistent with that found for the island-wide population of HM landslides (Hughes & Schulz, 2020). The Quebrada clay loam series dominates LLW with Humatas and Montegrande clay series playing a secondary role. However, about two-thirds of the mobilized sediments originated from areas with Humatas clay soils with significant contributions from Caguabo and Agüeybana clay loam series (43% and 26%, respectively; Figures 7d–7f). The Caguabo, Agüeybana, and Humatas series all displayed a high propensity to landslides with FR values greater than 1. About three-fourths of LLW is covered by forests from where ∼77% of the mobilized sediment originated. In contrast, only 8% of LLW is covered by cropland that mobilized ∼15% of total sediment (Figures 7g–7i). Cropland displayed a very high susceptibility to landslides with an overall FR value of 1.8, similarly to that found for all HM landslides throughout PR (Hughes & Schulz, 2020; Ramos-Scharrón et al., 2020). Surprisingly, forested areas displayed an overall FR of almost 1, which suggests a moderate level of landslide susceptibility. The effects of slope appear to be consistent for all land cover classes as areas with slopes greater than 30° tended to have FR values greater than 1 for all types (Figure 8a).

Details are in the caption following the image

Factors potentially controlling landslide sediment mobilization in LLW. Proportion of LLW surface area (%LLW), scar volume, and associated frequency ratios (FRs) in relation to geologic terrane (a–c), soils (d–f), and land cover (g–i). Error bars in (c, f, and i) represent the 95% confidence intervals in FR calculations.

Details are in the caption following the image

Combined effects of land cover, slope, and distance to roads on landslide scar area using joint frequency ratios (JFRs). JRFs for the combined effect of land cover and slope (a), land cover and distance to road (b), and slope and distance to road (c). Blank areas represent conditions not found within LLW. LLW, Lago Lucchetti Watershed.

LLW has 909 km of roads for an overall road density of 20.7 km km−2 (Figure 9). This road density is similar to some of the highest reported in the literature, and that is for areas where timber harvesting zones are accessed by a combination of haul roads and skid-trails (e.g., ∼23 km km−2 in Malaysia; Sidle, Sasaki, et al., 2004). In LLW, primary and secondary public roads represent only 5% of the road network. The remaining 862 km are mostly private access and service farm roads, out of which only 78 km are paved and 831 km are unpaved. Based on the 2010 land cover map, we estimate that 129 km (15%) of the farm roads are within active farms while the remaining 33 km are within abandoned farms.

Details are in the caption following the image

Roads in LLW: (a) watershed-scale view of the distribution of roads within LLW; (b) oblique aerial image taken from a drone of one of the coffee farms located within LLW [photo of Hacienda Candelaria by (c) J. Cruz Quiñones, Protectores de Cuencas Inc.]; and (c) close-up, hillshade view of a portion of LLW showing the location of landslides and the road network. LLW, Lago Lucchetti Watershed.

The proportion of LLW that lies in close proximity to roads is high. About 90% of all land within LLW is located within 50 m of a road and there is no area further than 306 m from any road (Figure 6m). Four metrics prove that roads and slopes play a vital role in the propensity of LLW to landslides. First, is that about 77% of all HM mobilized sediment occurred within 5 m from roads (Figure 6n). Second, is that FR values are significantly greater than 1 for areas within 5 m of roads (Figure 6o). Third, JFRs for both cropland- and forest-covered areas are above 1 for areas within 30  and 20 m from roads, respectively (Figure 8b). All active cropland is within 30 m roads making it highly susceptible for landslide mobilization. Forest-covered portions of LLW within 20 m of roads make up 49% of the entire watershed or 63% of all forested lands. Most of the forested areas in close proximity to roads are interpreted here to represent abandoned cropland. The high JFR values for shrub-covered land farther than 110 m from roads are insignificant as these areas represent only 0.3% of LLW (Figure 8b). Fourth, is that surfaces with slopes ranging from 30° to 70° and within 50 m of roads consistently displayed JFR values well above unity (Figure 8c). Roads are known to induce instability by oversteepening slopes and by focalizing runoff onto steep portions of the landscape which enhances soil saturation (Montgomery, 1994; Sidle & Ziegler, 2012; Vuillez et al., 2018).

HM mobilized 315,540 Mg of sediment from within 50 m of roads which leads to an area-normalized value of 8,600 Mg km−2. Although this is small compared to the highest road-related landslide mobilization rates found in the literature (278,000–4,823,500 Mg km−2 yr−1 from China; Sidle, Furuichi, & Kono, 2011; Sidle, Ghestem, & Stokes, 2014), this is significant for PR. In a mostly forested area in eastern PR with a land use history that does not include coffee farming, landsliding rates within 85 m of roads are ∼250 Mg km−2 y−1 (Larsen & Parks, 1997). Therefore, mobilization of sediment by HM landslides within 50 m of roads in LLW represents the equivalent of ∼34 years of landslides in the eastern part of the island.

The effects of roads in mobilizing sediment via landsliding are consistent for cropland and forest-covered areas. All of the 42,220 m3 of sediment mobilized within croplands occurred within 50 m of a road (Figure 10). This is almost inevitable as road densities within active cropland averages ∼36 km km−2 in LLW. The effects of roads in forested lands are not as abrupt as in cropland, yet 86% of all mobilized sediment within forests (194,500 m3 or 93% of the total) occurred within 50 m of roads. Although a process-based explanation of how roads incite such a large magnitude of sediment mobilization by landslides cannot be discerned with the methods employed here, it is undeniable that roads correlate to a high level of slope instability. Roads within what is presently covered by secondary forests are remnants of abandoned coffee farm transportation networks. Therefore, landscape alterations associated with road construction for past and present agricultural activity are having a lasting effect on LLW that facilitates a large amount of sediment mobilization by landsliding. This is similar to what has been found elsewhere for abandoned agricultural ditches and terraces (e.g., Brandolini et al., 2018; Persichillo et al., 2018). Therefore, roads in abandoned farms represent a type of anthropogenic memory (a la Brierley, 2010) that needs to be considered if sediment yields in the LLW are to be curtailed.

Details are in the caption following the image

Combined effect of land cover and distance from roads in landslide mobilization. (a) Summed volume of sediment mobilized relative to distance from roads for areas with varying land covers. (b) Proportion of mobilized sediment volume per land cover type relative to distance from roads.

4 Conclusions

The 2,318 shallow landslides triggered by Hurricane María within the 43.8 km2 LLW in western Puerto Rico mobilized from 230,510 to 430,330 Mg of sediment for an equivalent basin average vertical denudation of 4.4–8.3 mm. About 43% of the mobilized sediment is estimated to have been delivered to the stream network during the hurricane, which is equivalent to 4–100 years of surface erosion from 8%–30% of the watershed currently under active cultivation. The majority of the landslide-mobilized sediment during this storm was released from 30° to 70° hillslopes within 50 m of roads. Landslides could be the dominant source of sediment from this watershed if Hurricane María type events occur at least once every 3–8 years. This seems feasible given that maximum 1-h and daily rainfall rates for this storm represented intensities of only a 1- to 10-year recurrence interval for the study watershed. The fact that such a relatively common rain event for this part of the island led to such a significant sediment mobilization is likely the product of the high vulnerability of the landscape to slope failures and the relatively large portion of the watershed exceeding landslide-triggering rainfall thresholds during the hurricane. In this particular case, the high propensity for landsliding is strongly linked to the high road densities associated with active and, more importantly, presently abandoned coffee farms. The findings of the study have repercussions on watershed management strategies for Puerto Rico and other similar roaded, tropical, and subtropical steep landscapes.


No external funding was solicited nor required for the completion of this research. B. Vest’s involvement was covered through faculty startup funds provided by the College of Liberal Arts at UT-Austin. Our sincere gratitude to T. Beach and three anonymous reviewers for their thorough comments on earlier versions of this manuscript.

    Data Availability Statement

    All geospatial data used for this article are archived at Zenodo (http://doi.org/10.5281/zenodo.4379143).