Rapid assessment of urban mega-gully and landslide events with Structure-from-Motion techniques validates link to water resources infrastructure failures

Mega-gullies and landslides pose significant hazards to urban development on steep terrain. Water resources 15 infrastructure failures (WRIFs), such as leaks and breaks in water supply pipes, have been postulated as a trigger of mass movement events but data for validation has been challenging to acquire since earthwork proceeds quickly after events to repair roads and other infrastructure. Urban development in Tijuana, Mexico was monitored for a five-year period to document the occurrence of mega-gullies and landslides, including sediment volumes. A rapid assessment approach was developed based on photogrammetric observations from an unmanned aerial vehicle (UAV) and Structure from Motion (SfM) digital 20 processing. Three hazardous mass-movement events were observed including two mega-gullies and one landslide. Furthermore, all three events were linked to WRIFs. Frequency analysis points to the annual probability of a WRIF-based erosion event in the range of 40-60%, which is far higher than design levels typically used for urban stormwater infrastructure (5-10%). Additionally, sediment modelling points to WRIF-based erosion as a non-negligible contributor to sediment generation. These results suggest that WRIFs are a significant contributor to erosion hazards facing urban development on 25 steep terrain, and call for expanded monitoring to characterize the occurrence and modes of WRIF-based erosion events.

. Mega-gully and landslide hazards are increasing at a time of rapid urbanization as a result of limited oversight of planning and construction as well as socio-economic pressures that force populations to settle in highhazard areas (Hardoy et al., 2013;Retief et al., 2016;Miller et al., 2019). Moreover, mass movement hazards are concentrated in developing middle-and low-income regions of the world where unregulated settlement occurs on steep slopes (Anderson 35 communication, 2016) have observed that WRIFs trigger mega-gullies and landslides that pose life-threatening hazards. The objectives of this paper are three-fold: (1) to report the effectiveness of SfM photogrammetric techniques for rapid erosion assessment following WRIF events, (2) to document the frequency and magnitude of sediment mobilization from WRIF events, and (3) to evaluate the significance of WRIF events with respect to mass movement hazards and sediment budgets at neighborhood-and watershed-scales. 70 The remainder of the paper is organized as follows: Section 2 (Methods) presents a site description, SfM-based observational methods, and watershed modeling methods to sediment budgets; Section 3 (Results) presents data on the frequency and magnitude of WRIF-based mass movement and neighborhood-and a comparison of WRIF-based fluxes to other mechanisms of sediment generation. The paper concludes with discussion (Section 4) and conclusions (Section 5).

Site Description
The Los Laureles Canyon Watershed (LLCW) is an urbanizing binational catchment that overlaps the western-most portion of the USA -Mexico border. LLCW flows from the city of Tijuana, Mexico, into the Tijuana River Estuarine Reserve, USA ( Fig. 1). Sediment from excess erosion and solid waste transported in the channel network from Tijuana have impaired the estuarine ecosystem (Weis et al., 2001). low in the southern part of the watershed and in the areas with infrastructure failure as evidenced by a high marginality index and a low fraction of homes with piped water or drainage as compared with other areas of Tijuana (Biggs et al., 2014).

Study Design 95
Hydrologic conditions, slope instabilities, and sediment generation rates were monitored in the LLCW for a 5-year period beginning in January 2013 and ending in April 2018. To measure rainfall, a tipping-bucket rain gauge station ("LLCW raingage" in Fig. 1) was installed in the watershed, and to measure stream stage, a pressure transducer (PT) (Solinst, water level logger) was installed in a concrete channel at the watershed outlet and logged conditions at 5-minute intervals (Fig. 1).
Upon detection of flow at the watershed outlet, field personnel travelled to the site, performed a visual inspection of site 100 conditions, and, upon observation of mega-gullies and landslides, collected two types of data about the WRIF erosional features: (1) photogrammetric surveys (RGB images) were performed with a nonmetric camera (GoPro Hero3+) from either a ground-based or aerial platform, and (2) ground control points (GCPs) were acquired by differential GPS (Magellan Pro Mart 3) with sub-centimeter to 5 cm accuracy (Magellan Systems Corporation, San Dimas, USA). These primary data were used to create Digital Surface Models (DSMs), and in turn, estimates of sediment volumes and their impacts on sediment budget, 105 as well as to document safety hazards to the people living in the watershed and downstream ecosystems.
A long-term record of rainfall is available from the NOAA Tijuana River Estuary gauging station, located near the outlet of the LLCW, which provides daily rainfall for the period 1980 to 2018. These data are used here to estimate the return period of storm events during the study period. Data from the tipping-bucket rain gauge (LLCW Raingage in Fig. 1) were used to force a watershed erosion model, validated with stream gauge data and observed sediment loads at the outlet. The model was 110 used to compare sediment generation from WRIF features with the other watershed processes described in section 2.4.

Image acquisition and processing
Photogrammetric surveys using a nonmetric camera (GoPro Hero3+) were completed using either an Unmanned Aerial System (UAS) (DJI, Phantom2) or a telescoping painter's pole (approximately 2-3 m long). The UAS is advantageous for relatively large and wide erosional features compared with the painter's pole, which can better access relatively small, narrow, and deep 115 erosional features (Gudino-Elizondo et al., 2018a, Taniguchi et al., 2018. Images were acquired once per second using the time-lapse capture mode from different angles to ensure a high overlap between photographs and to reduce the shade in each image (Castillo et al., 2015) and doming deformations (James and Robson, 2014).
The sediment volume mobilized was estimated using a four-step procedure: (1) Imagery were combined with a subset of the GCPs to calibrate the camera and produce Structure from Motion (SfM) point clouds following general workflows (Agisoft 120 LCC, Russia, Version 1.4.4), (2) SfM point clouds were converted to a digital surface model (DSM) (Agisoft LCC, Russia, Version 1.4.4), (3) erosional volumes were computed (ArcGIS 10.6.1, ESRI, Redlands, California) by subtracting the DSM from a reference DSM representative of the pre-event land surface (Wheaton et al., 2010), and (4)  integrated to calculate the total sediment volume. Volumes were converted to mass using a bulk density of 1,600 kg /m 3 corresponding to very fine sand (USDA, 2018). 125 Pre-event topography was based on either a 2014 aerial LIDAR survey (1 m resolution DSM with a 0.11 m vertical RMSE, NOAA, 2014), or with UAS-based DSMs generated with imagery collected before the failure event (Table 1). The horizontal and vertical RMSE of the DSM, or geo-registration error, was estimated using the subset of the GCPs not used to produce the SfM point cloud, called Error Control Points (ECPs). Previous work indicates that 4 to 5 GCPs with a few additional ECPs are adequate for SfM processing (James et al., 2017). The RMSE for the DSM creation was computed as the square root of the 130 sum of the squared errors for each ECP (Alfonso-Torreño et al., 2019). in the field and compared to length estimates from the SfM point cloud as described in Gudino-Elizondo et al. (2018a).
Additionally, pre-and post-event ground elevations were compared along transects outside the disturbed region where no topographic change was observed.

Watershed Modeling
The Annualized AGricultural Non-Point Source (AnnAGNPS) model (Bingner et al., 2015) was applied to the LLCW to 140 simulate discharge and sediment load during storm events and to develop an inventory of sediment generation rates by mechanism at the watershed scale. The AnnAGNPS model was previously calibrated and validated for runoff and observations of sediment generation in LLCW (Gudino-Elizondo et al., 2018a, 2018b, 2019b, and the applications here rely on this Measurements and modeling supported an inventory of sediment generation and load from four mechanisms: (1) sheet and rill erosion, (2) gully erosion, (3) channel erosion, and (4) erosion from WRIF. Sediment generation was considered as the total mass of sediment mobilized, while the sediment load was the quantity of sediment observed at the watershed outlet. Sediment 150 load from WRIF was calculated by multiplying the erosion volume per event times the Sediment Delivery Ratio (SDR). For mega-gullies, the SDR was set to 1 based on field observations and modeling work described in Gudino-Elizondo et al. (2019b). Conversely, the SDR was set to zero for the landslide based on field observations that displaced sediment was intercepted by the road network and mechanically removed or repositioned on the hillslope (Vigiak et al., 2012). Of course, subsequent rainfall events may cause the repositioned sediment to be later mobilized and moved towards the stream network, 155 so our estimates of load correspond only to the period of observation.

Hazard assessment from water resources infrastructure failure
Reports of the damage caused from mega-gullies and landslide were compiled from residents and local agencies. Primary data from the three events are described, including impacts to transportation, housing, urban infrastructure and downstream ecosystems and communities in the study watershed. The specific soil loss (SSL) of the WRIF mega-gullies was calculated as 160 the total erosion (m 3 ) normalized by the watershed area (m 2 ) and was then compared to the observed SSL in the study watershed and to other studies reported in the literature. A detailed description of safety hazards and the contribution to the total sediment budget of each WRIF event is described in section 3.

Results
A total of 14 storm events were observed during the 5-year study period, based on a flow threshold of 1 m 3 s -1 (or ~15 cm of 165 water in the channel) at the gaging station, which corresponds to a depth of rainfall ranging from 6.5 to 13 mm. The total depth of the 14 storms was 322 mm, or 35% of the total rainfall (907 mm, 181 mm y -1 ) for the 5-year period. Mass movement from WRIFs were observed during three of these events, each characterized by a 1-2-year return period. WRIFs leading to mass movement were not observed between storm events or during smaller storm events (< 23 mm).

SfM surveys and social impacts by event 170
Three storm events caused WRIFs that led to hazardous mass movements. One event involved a landslide and two involved formation of mega-gullies. The detailed observations of each feature are described in sections 3.1.1 to 3.1.3.

Landslide
A large landslide occurred during a storm event on 15 May 2015 (Fig. 2). More than 20 houses were damaged affecting more than 100 people (Fig. 2b, 2c). Based on the daily rainfall total (23 mm) and the long-term rainfall record at the NOAA Tijuana 175 River Estuary Gage, the return period of the storm is 1 year. The landslide was attributed to a WRIF based on resident reports that seepage from the slope along with incipient cracks were observed for several days immediately before the failure incident.
This observation led to the evacuation of the residents when the evolution of the cracks was evident. The infrastructure failure wetted the soil, and the landslide was then triggered by the rain event. Broken water mains were also observed after the landslide (Fig. 2c,   vertical errors estimated here. Elevation differences outside of the disturbed area were 0-7 cm. The mean difference between measured and modeled lengths of objects at the site (e.g., sewer manhole covers) was less than 3 cm. These different methods 195 all suggest the error was less than 7 cm and within the range expected for the observation distance.

Mega-gully A
A mega-gully formed along an unpaved road following a storm event on 15 September 2015. Based on the daily rainfall total (31 mm) and the long-term rainfall record at the NOAA rain gauge, the return period of the storm was 1-2 years. This megagully was attributed to a WRIF based on resident reports that leakage from a broken pipe was observed upstream immediately 200 after the failure event (personal communication, Tijuana Metropolitan Planning Institute). In this case, erosion caused by the storm event undermined the water main, which subsequently broke and enlarged the gully as a result of high velocity water jets from the pressurized water main. The mega-gully was 98 m long, with a maximum width of 8 m and maximum depth of 4 m (Fig. 3). The generated sediment mass was estimated as 1,360 ± 65 tons. December2015-08 August 2016) before the road was repaired. 215

Mega-gully B
A second mega-gully formed along an unpaved road following a storm event on 16 December 2016. Based on the daily rainfall total (33 mm) and the long-term rainfall record at the NOAA rain gauge, the return period of the storm is 1-2 years. The megagully was largest at the upslope position and decreased in cross sectional area with distance from the broken pipeline. The mega-gully was 202 m long, with a maximum width of 10 m and maximum depth of 7 m. Imagery was collected for the SfM 220 processing using a telescoping painter's pole (Fig. 4d), and sediment generation was estimated to be 4,340 ± 408 tons. The RMSE (ECPs, n=10) of the DSM obtained from SfM was 3.5 cm in the horizontal and 5 cm in the vertical. Elevation differences outside of the disturbed area were 0-5 cm, which is consistent with the accuracy of the method. For example, differences between measured and modelled lengths of not-deforming objects at the site (e.g., water supply pipes shown as 230 blue stars in Fig. 4c and Fig. 4d) were less than 1 cm.
For this second mega-gully event, mass movement was again triggered by erosion that undermined the water main, which subsequently broke and enlarged the gully by discharging piped water directly onto the hillslope. Broken water main pipes were noted during the rapid-response survey ( Fig. 4a and 4c). The mega-gully also impacted public transportation and life quality in the neighbourhood for 6 months (based on Google Earth imagery) and interrupted water supply for 1 month, affecting 235 more than 200 people (personal communication, Tijuana Metropolitan Planning Institute).

Comparison of Sediment Generation Sources
Application of the calibrated AnnAGNPS watershed model to storm events for 2012-2017 yielded daily estimates of rainfallbased sediment generation by sheet and rill erosion, gully erosion and channel erosion. Table 2

*Sediment quantified using field measurements, **Sediment quantified using the AGNPS model. 245
This analysis shows that mass movement associated with WRIFs was significant on an event basis. Mega-gully B generated 4,340 tons (Table 2), which is approximately 80 times the area-normalized annual erosion rate for gullies (tons/ha) and 10 times the total sediment generated by other rainfall-generated gullies (Gudino Elizondo et al., 2018a, Gudino Elizondo et al., 2018b. The WRIF-triggered landslide mobilized more sediment than all of the rainfall-based processes combined, while the 250 mega-gullies triggered by pipe failures and hydraulic mining were responsible for 16 and 20% of the total sediment generation across the watershed (Fig. 5). The proportion of sediment generated by each erosional process differed markedly between the landslide event and the two mega-gully events (Fig. 5); rainfall-generated gullies contributed more sediment during the landslide event because peak discharge, the main control on gully formation, was higher during the landslide storm event (19.5 cms at the outlet) than during 255 the two mega-gully events (~5 cms) (Gudino-Elizondo et al., 2019b).

260
The total sediment generation and load were computed for the 5-year study period by integrating over all storm events (Table   3). On a five-year basis, WRIFs contributed 5% of the total sediment generation and approximately 2% of the total sediment load at the watershed scale. While the sample size here is small, the frequency of WRIF-based erosional events can be estimated in several ways: three hazard events occurred over a period that had 14 rainfall events (21% of rainfall events), two out of five years had at least one hazard event (40% chance per year), or three events occurred in five years (60% chance per 265 year). The small sample size implies a high degree of uncertainty in all of these estimates; nevertheless, these rates of occurrence are far higher than typical design standards for water resources infrastructure in urban areas. For example, large flood control channels are typically designed with a 0.2-2% annual exceedance probability, and smaller drainage systems in urban areas are often designed for 5-10% annual exceedance probability. Hence, WRIF-based hazards observed during this study are many times more frequent (21-60%) than typical design standards for flood control systems in urban areas (0.2-10%) 270 and thus deserving of greater attention for public safety, infrastructure resilience and environmental protection.

Comparison to Previous Observations 280
The mega-gullies observed here are large compared to rainfall-generated gullies surveyed in the study area (Gudino-Elizondo et al., 2018a), which had a mean gully width of 1.5 m and a mean depth of 0.5 m. In contrast, mega-gully A is up to 8 m wide and 4 m deep, and mega-gully B is up to 10 m wide and 7 m deep. Mega-gullies A and B are also long with lengths of 100 and 200 m, respectively. Mega-gullies A and B were also more developed than rainfall-generated gullies, with greater connectivity to the stream channel, which enhances sediment delivery to the stream network. 285 Figure 6 shows that the specific soil loss (SSL, the average depth of soil loss in the watershed) from mega-gully B was exceptionally high compared to rainfall-runoff gullies in the LLCW (Gudino-Elizondo et al., 2018a) and compared to sites reported by Castillo and Gómez (2016), which included sites spanning different land uses and precipitation regimes. The SSL from mega-gully A was comparable to other gullies observed in the study watershed, which has higher rates of SSL than the set of sites reported by Castillo and Gómez (2016).  The landslide caused by the WRIF was the single largest erosional feature observed in the watershed during the study period.
Landslides occur throughout Tijuana, with more than 40 landslides from 1992-2012, including a landslide that damaged 19 buildings (Oliva-González et al., 2014), which is comparable to the LLCW slide (20 buildings damaged). However, data 300 describing the sediment displaced by these 40 landslides and potential connections to WRIFs were not available.

Rapid methods for monitoring erosional features
Landslides and mega-gullies have complex topographies and are poorly suited to the application of traditional surveying techniques such as total stations, but are well suited to photogrammetric characterization using SfM. In our study, 305 photogrammetric data were effectively captured for mega-gullies roughly 5-10 m wide, 5-10 m deep, and >100 m long using a camera mounted on a telescoping painter's pole, and landslides were safely characterized using a UAS-based platform. The presence of urban infrastructure in photographs (e.g., concrete pads, sewer structures) also presented opportunities for ground control points and accuracy checks. Accuracies achieved from these observations (i.e., 3 cm horizontal RMSE and 5 cm vertical terms of point density and data accuracy, but they are typically quite heavy compared to cameras, require more sophisticated 315 spatial referencing systems (e.g., inertial navigation units), often present occlusion artifacts, and are expensive (Izumida et al., 2017;Mazzoleni et al., 2020). Our study demonstrates that both ground-based and UAS-based photogrammetry allow for rapid documentation of hazardous erosional features with minimal equipment and low labor requirements.

WRIFs and the sediment budget
Stochasticity in WRIFs and WRIF-based sediment hazards is high. Failures may or may not happen in any given storm (here 320 we observed 3 failures in 14 storm events), and when failures occur, the volume of sediment generated across three events varied by over an order of magnitude. This makes it difficult to generalize and estimate sediment generation by infrastructure failure for other events lacking field observations. However, the data do allow a first-order estimate of annual-average sediment generation from WRIFs which is useful for sizing sediment basins that protect downstream ecosystems from excess sedimentation and for estimating average-annual excavation costs. Such estimates would not likely be applicable outside of 325 the LLCW, but the photogrammetric methods deployed here to monitor sediment generation are easily transferrable to other systems, and data on sediment generation from multiple sites would provide a basis for improved understanding and possibly transferrable models.

Feedbacks between erosion and slope instabilities from WRIFs: opportunities for hazard mitigation
Erosion and hazards produced by WRIFs were either exacerbated or triggered by erosion during storm events. The observed 330 landslide was triggered by a storm event, but the event was preceded by the water main leak. The observed mega-gullies formed after local runoff initially undermined water mains, which then broke and discharged water onto the hillside, triggering more severe gully erosion. This suggests that WRIFs, storm events, and slope instabilities are interdependent. Moreover, this opens the possibility of reducing mass movement hazards through improved design, management and oversight of water resources infrastructure. Whereas rapid urbanization is broadly linked to minimal levels of governance and institutional 335 oversight of urban infrastructure, especially in least-developed countries (Borelli et al., 2018), water resources infrastructure benefit from relatively high levels of planning, design, engineering and oversight (Whittington et al., 2009;Cook, 2011). For example, mass movement hazards could be reduced by aligning water mains away from topographic low spots susceptible to gully formation, and away from hillslopes that may be susceptible to creeping displacements that stress pipes and cause leaks.
Pipeline specifications could also be changed to promote greater ductility, or resistance to failure, under hillslope displacement 340 (Honegger et al., 2010;Han et al., 2012). In turn, the water resources infrastructure would benefit from fewer leaks and breaks and higher levels of reliability.

Implications for hazard mitigation in other urban contexts
Landslides and mega-gullies like those observed in Tijuana have been reported across cities in middle and low-income countries where unregulated settlement occurs on steep hillslopes , but also in developed countries. For example, in the city of San Diego, California (USA), soil erosion caused by a storm on January 5, 2016 undermined a 30-foot section of sewer causing failure and prompting a spill of more than 6.7 million gallons of untreated sewage that severely eroded the riverbank and negatively impacted downstream ecosystems (Garrick, 2020). Nevertheless, what is clear is that even though the sample size of events reported in this analysis is small, the severity of the events involving WRIFs is high. Housing, transportation, and utilities that serve hundreds of people living in the watershed are impacted by WRIFs in Tijuana. WRIF-350 based mass movements also contributed a significant amount of sediment to the total watershed load, which negatively impacted habitat and aquatic ecosystems, and further increased downstream infrastructure maintenance costs (Brand et al., 2020). Acknowledging the challenges of monitoring, as addressed here, what becomes clear is a need for more widespread monitoring of landslides and mega-gullies and documentation regarding the role of WRIFs. It is possible that a substantial fraction of the most hazardous mass movement events in cities are linked to WRIFs, and that significant hazard reduction can 355 be realized by addressing WRIFs.

Conclusions
Urban development in Tijuana, Mexico was monitored for a five-year period to document the occurrence of mega-gullies and landslides, including sediment volumes. Over a five-year period with 14 storm events, two mega-gullies and one landslide were observedand each occurring during a rainfall event. While the link between rainfall and erosion hazards is well known, 360 monitoring showed that all three events were associated with a Water Resources Infrastructure Failure (WRIF). Mega-gullies occurred after a break in a water supply pipe, which unleashed a highly erosive, high-velocity water jet. Moreover, pipe breaks were observed to occur after rainfall and runoff formed a gully that undermined structural support for the water supply pipe.
Hence, mega-gullies formed from a two-step process: (1) a water supply pipe breaks after the formation of a gully, and (2) a mega-gully is formed from the high velocity jet out of the water supply pipe. The observed landslide was also linked to a two-365 step process: (1) a water main leak saturated the hillslope creating the preconditions for a landslide, and (2) heavy rainfall triggered the landslide. Erosional features caused by WRIFs were larger than features generated by local rainfall and runoff, produced a significant amount of sediment on an event basis, and presented major safety hazards to downstream communities and ecosystems at the neighborhood and watershed scale. The limited data suggest that WRIF-based erosion events occur with an annual frequency of 40-60%, which is far higher than typical design standards for stormwater infrastructure (5-10% annual 370 exceedance probability). Modeling shows that WRIFs contribute, on average, 5% of the total annual sediment generation at the watershed scale, and up to 58% on a storm-event basis. Additional research is needed to improve estimates of the spatial and temporal frequency of erosional features caused by WRIFs, and to understand the significance of WRIF hazards at other spatial and temporal scales and in other geographic contexts. Furthermore, the hazards posed by WRIFs within development on steep terrain calls for greater attention to infrastructure design and maintenance.
Structure from Motion (SfM) photogrammetric techniques helped to rapidly and safety assess the volume and shape of megagullies and landslides. Using imagery collected by either Unmanned Aerial Systems (UASs) or a camera on a hand-held pole, SfM techniques registered Digital Elevation Models (DEMs) with errors of ~3 cm horizontal RMSE and ~5 cm vertical RMSE which are in line with the needs for sediment budget applications. 380

Authors contribution
NG undertook data acquisition, processing and interpretation of the data, and prepared the manuscript with contributions from all co-authors. BS, TB and AG designed the research, and RB provided valuable guidance on the soil erosion modelling.