*3.4. Sediment Budget*

The annual trap efficiency varied from 0.79 to 0.98, and was 0.89 for the cumulative mass removed over 2006–2012 [33]. Uncertainty in sediment trap efficiency calculation, in particular the gradual filling of the traps during the year and subsequent decrease in trap efficiency, may have caused underestimation of sediment load at the traps (Table 6). Total annual sediment accumulation in the traps correlates with annual precipitation at the SDBF station (Figure 9).

**Figure 9.** Total sediment removed from the Los Laureles Canyon traps in the United States versus total annual precipitation between removal events from 2005 to 2012. The uncorrected amount of sediment removed is in black and the load corrected for trap efficiency is in grey. Annual precipitation is from the San Diego Brownfield station. Modified from Biggs et al. [33].

**Table 6.** Simulated annual sediment yield and total observed (in tons) at the watershed outlet by the erosion process. The range values for channel contribution and total yield assumes channel erosion is 25% (minimum value) and 40% (maximum) of the total simulated results from AnnAGNPS.


Total modelled sediment load correlates with the sediment load observed at the sediment trap (Figure 10), with the following errors: (i) pBIAS25% = 8.1, RMSE25% = 24,115 t (41% of the mean) considering a channel erosion contribution of 25% of the total hill-slope sediment production, and (ii) pBIAS40% = 15.4, RMSE40% = 32,570 (55% of the mean) considering a channel erosion contribution of 40%. These model efficiencies were moderate-to-good according to other studies [44,49,52] reporting relatively similar values.

**Figure 10.** Time series of the relationship between observed and simulated annual sediment load at the LLCW outlet, assuming a channel erosion contribution of 25% and 40% of the total hill-slope sediment production.

The default values of <sup>τ</sup>*<sup>c</sup>* for conglomerate soil types (CfB and CbB) (τ*<sup>c</sup>* = 64 N·m−2) resulted, on average, in underestimation of the total sediment load at the outlet, so it was changed during calibration to <sup>τ</sup>*<sup>c</sup>* <sup>=</sup> 32 N·m−<sup>2</sup> to fit better with the observed sediment yield in the sediment traps. The <sup>τ</sup>*<sup>c</sup>* value set in the calibrated model corresponds to very coarse gravel [37], which was consistent with the observed particle sizes of gravelly and cobbled soils in the LLCW [33].

Precipitation correlates with simulated sediment production from sheet and rill erosion, gully erosion, and total sediment yield, while sediment production by sheet and rill erosion correlates more closely with rainfall than gully sediment production does (Figure 11). A minimum precipitation threshold (~25–35 mm) for gully initiation was reported by Gudino-Elizondo et al. [9], which is consistent with the significant contribution by gullies to the total sediment production for those storm events with precipitation greater than 25 mm (Figure 11).

**Figure 11.** Simulated sediment load by erosion processes in LLCW. The vertical dashed lines indicate the range of the rainfall threshold for gully erosion observed in the field during 2013 to 2018.

Simulated sheet and rill erosion was the dominant erosional processes within the LLCW (Figure 12), which was also reflected in the event-wise rainfall-sediment relationships (Figure 11), especially for larger events. Total sediment load at the sub-watershed scale (AnnAGNPS cells) was dominated by cells characterized by sandy soil types (Lf) on steep slopes, which show evidence of frequent rill and gully formation (Figure 12).

The observed sediment in the trap is finer (higher silt fraction) than both the hill-slope sediment and the AnnAGNPS-simulated sediment load (Figure 2). This suggests that either more sand is being retained in storage on the hill-slopes and in the channel than is simulated by the model, or that silt is preferentially eroded from soils that have a mixture of silt and sand, or that soils with high silt fraction contribute more to the load than is being modelled. The particle size in the Mexico sediment trap is coarser than the US sediment trap, which suggests either retention of sand in the channel downstream of the Mexico sediment trap, or high loads of silt from the sub-watershed outside of the Mexico sediment trap sub-watershed.

**Figure 12.** (**a**) Sub-watershed sediment yield by gully erosion. (**b**) Total sediment yield by sub-watershed within the Los Laureles Canyon watershed and (**c**) total sediment production by contributing the drainage area under the current conditions and road-paved scenario.

#### *3.5. Scenario Analysis*

Half of the simulated sediment load at the watershed scale is generated by only 23% of the total watershed area under current conditions (red lines in Figure 12). These cells are hotspots of sediment production and, pending validation of erosional severity with additional field observations, could be prioritized for management activities to reduce sediment production at the watershed scale.

The model scenario suggests that, on annual average, paving all the roads in the hotspots would reduce sediment production by 30% (Figure 12). However, storm-wise total runoff increases by an average of 10%, and peak runoff increases from 1.6% to 21% (Table 7). The projected peak discharge increased the most for the medium-sized events (40–49 mm, two-year recurrence interval), and not for the largest event (81 mm, 25-year recurrence interval), which suggests that paving roads in hotspots could be suitable for the study watershed without increasing peak discharge for the largest events. This may be the most responsible factor for flood damage.

#### **4. Discussion**

AnnAGNPS simulated total water and sediment load with satisfactory agreement with observed total event runoff and sediment yield in the LLCW. The simulated total runoff and peak discharge were more accurate for medium-sized events (event precipitation between 2 and 20 mm, Figure 7), which are the most frequent events in the region. The model generally underestimated peak discharge for small storms and overestimated peak discharge for large storms, which is partly due to underestimation of the peak rainfall rates of small storms and overestimation of the peak rainfall rates of large storms by the SCS Storm type model [33]. Further research on the impact of paving on peak discharge for a range of storm sizes and sequences is needed.


**Table 7.** Rainfall and simulated increases in peak and total discharge volume at the outlet under current and scenario (hotspot paving) conditions for the 14 largest storm events.

The AnnAGNPS model satisfactorily simulated ephemeral gully erosion rates in the LLCW at the neighborhood scale [14], which helped identify parameter values for use at the LLCW scale. The model performed well compared with other models applied in semi-arid environments [23,30,49,51,53] which supported its use for runoff and sediment budgets in this watershed. However, uncertainties in soil-resistance-to-erosion parameters, especially critical shear stress for cobbled soils, may affect sediment production by gully erosion. This suggests that more field and laboratory data are necessary to have more accurate sediment yield estimates at the watershed scale.

The SSC from 10 grab samples (Figure 8c) collected during the largest storm event correlated with modelled SSC, which suggests the model represented spatial variations in sediment production within the watershed. Modelled erosion was sensitive to the fraction of highly-erodible Lf soil type, which generated 61% of the total sediment load. Most of the AnnAGNPS cells that contribute significantly to the total sediment load (hotspots) had both highly erodible soils (Lf) and steep slopes (>30%) that encourage gully sediment production (Figure 12). No SSC samples were available for areas drained by cobbled and gravel soils in the northern part of the watershed, so future work should include more grab samples from sub-watersheds draining cobbled soils.

The RMSE of the AnnAGNPS model for sediment load was 41% and 55% of the mean value considering 25% and 40% of channel erosion contribution, respectively. Our observed values of sediment load at the outlet likely underestimate the total load because our method for calculating trap efficiency does not account for a reduction in trap efficiency as the trap fills during the wet season. This underestimate is likely largest during wet years when the trap is full at the end of the season,

so future research will explore the impact of reduction in available trap capacity on trap efficiency and estimated sediment load. Channel evolution is also not well characterized by the AnnAGNPS model, which reduces the performance of the model to simulate the observed behavior of the system. Taniguchi et al. [38] noted that urbanization caused extreme channel enlargement in the LLCW, which suggests the necessity to implement and couple a more sophisticated channel evolution model to AnnAGNPS such as the channel evolution computer model (CONCEPTS) [54,55] to better simulate the sediment production at the LLCW scale.

Simulated gully erosion represented approximately 16% and 37% of total sediment production considering 25% and 40% of channel erosion contribution, respectively. This was relatively close to other estimates for human-disturbed watersheds. Bingner et al. [56] reported that ephemeral gullies were the primary source of sediment (73% of the total) in agricultural settings within the Maumee River basin, USA. De Santiesteban et al. [57] found that ephemeral gullies contributed 66% to total soil loss in a small agricultural watershed. Taguas et al. [23] found that contribution of gully erosion to the total soil loss varies substantially depending on the management, on average, from 19% to 46% under spontaneous grass cover and under conventional tillage management. Previous studies reported gully erosion in agricultural and partially urbanized watersheds even though gully erosion can be a significant, and often neglected, portion of the sediment budget. We likely underestimate gully contribution since the model assumes gullies on roads are filled only once per year, while field observations suggest that main roads are repaired several times per year, after every storm that generates gullies. Our estimates of gully contribution are also sensitive to the mapped distribution of fine-textured soils that generate most of the gullies. Future research could refine the soils map and test for the sensitivity of gully filling frequency on the gully contribution.

Our estimate of the contribution of channel erosion to the sediment load (25%–40%) is smaller than other studies, which report channel contributions ranging from 67% [58] to 85% [59] of the total sediment yield in urban areas. The relatively large contribution from hill-slope sources in our study area is likely due to persistent soil exposure and erosion, including vacant lots and unpaved roads, that characterizes urbanization in Tijuana [8] and possibly other cities in developing countries.

In our watershed study, 50% of the sediment production under current conditions is generated only from 23% of the watershed area. Paving all the current unpaved roads in these "hot spots" would reduce sediment production by 30% compared to the current conditions, but it would also increase total discharge by 2%–17% and peak discharge by 2%–21%. The smallest increase in peak was for the largest event, which suggests that the impacts of paving may be small for the events that cause the most flood damage. This is consistent with other studies that document proportionately large impacts of urbanization on the smallest events, and declining impact for larger events [60], even though more complete documentation of the impact of paving for a range of storm sizes under different antecedent moisture conditions is necessary.

This investigation highlights the necessity to implement management activities to mitigate soil erosion such as stabilization of unpaved roads and other management activities (i.e., revegetation, sediment basins, channel stabilization, etc). Future studies should evaluate the uncertainty of the model-estimated parameters as well as implications in scenario analysis [14], which are critical for proper sediment management in the LLCW and potentially in other rural urbanizing watersheds, particularly those in developing countries. Our study highlights the relative importance of various erosion processes, and also key uncertainties for future investigation.
