Comparison of Two Watershed Models for Addressing Stakeholder Flood Mitigation Strategies: Case Study of Hurricane Alex in Monterrey, México

The study region is the Santa Catarina watershed (SCW) ($\mathrm{1,831}{\mathrm{km}}^2$) in northeastern México, which encompasses Monterrey, capital of Nuevo León [Fig. 2(a)]. A subtropical, semiarid climate characterizes the watershed (Návar and Synnott 2000), which is a subbasin of the San Juan River ($\mathrm{33,538}{\mathrm{km}}^2$) that provides water to Monterrey through reservoir operations at El Cuchillo Dam (Scott et al. 2007). The metropolitan area occupies the lower areas of the SCW and has a population of nearly 4 million, with an annual growth rate of 1.3% from 1990 to 2010 (INEGI 2011). As an important city in México (INEGI 2009), Monterrey has a diversified economy that supports a sprawling amount of urban land cover on relatively mild slopes. In contrast, the upper reaches of the SCW are predominantly rural areas because of the highly-sloped mountain ranges [Fig. 2(b)] that form part of the 1,000 km long Sierra Madre Oriental in northeastern México (e.g., Maqueda et al. 2008; Ferriño-Fierro et al. 2010). Table 1 lists the areal coverage of soil classes from the Food and Agricultural Organization (FAO) and land cover classes in the SCW, with the majority of the urban area and human settlements occurring in the lower reaches of the basin [Fig. 2(c)]. While only a small fraction of the SCW is used for agriculture, downstream areas in the San Juan River support important activities in this sector that are also subject to flood hazards (Scott et al. 2007). The primary land covers in the upper reaches of the SCW are submountainous shrublands and secondary shrublands at low to midelevations, and mixed woodlands on the higher mountain slopes. The sequence of folded ridges in the SCW leads to a trellis-like stream network consisting of ephemeral channels that increases flood hazards upstream of the city during storm events (Ramírez 2010).

Recurring floods from hurricane landfalls prompted the construction of a flood control structure at Corral de Palmas, known more commonly as Rompepicos Dam [Fig. 2(b)]. The dam has a gravity curtain composed of roller-compacted concrete, an elevation of 70 m and a length of 240 m (Ramírez 2011). The structure has two outlet structures: a secondary rectangular opening of 6 × 6 m at the base of the dam with a maximum capacity of $838{\mathrm{m}}^3/\mathrm{s}$ and a main Creager spillway with a crest length of 60 m and a capacity of $\mathrm{3,376}{\mathrm{m}}^3/\mathrm{s}$ at maximum water level (Ramírez 2011). Completed in 2004, Rompepicos Dam controls flooding generated in the uppermost reaches of the SCW and has served its design purposes during Hurricane Emily in 2005 and Hurricane Alex in 2010. For Hurricane Alex, only two hydrologic observations were available: a visually-estimated maximum water level behind Rompepicos Dam (2.5 m below the spillway or $650{\mathrm{m}}^3/\mathrm{s}$) and the continuous discharge record at the basin outlet (Cadereyta Station) by the Comisión Nacional de Agua (CONAGUA). However, both of these observations are considered to be uncertain given the flood damages occurring along the measured river reach at the Cadereyta Station and the visual inspection of the maximum level in the reservoir rather than a measured water depth. All other streamflow observations were compromised during the flood.

Hurricane Alex was the first tropical cyclone of 2010, having initiated in the Caribbean Sea and intensified in the Gulf of México to a Category 2 Hurricane before landfall in Tamaulipas, México (see Pasch 2010; Vitale et al. 2015 for event descriptions). Interactions of the tropical moisture with the orographic barriers in the Sierra Madre Oriental led to widespread rainfall in Nuevo León from June 28 to July 2 (Hernández and Bravo 2010). Large rainfall accumulations were observed in the sparse network of rain gauge sites near Monterrey [Fig. 2(b)], with 72-h totals ranging from 250 to 800 mm, and in some locations exceeding the mean annual precipitation in the semiarid region (SMN 2010). Ramírez (2010) estimated daily return periods between 20 and $>100\mathrm{years}$ at daily CONAGUA sites, though these tended to underestimate rainfall as compared to several automated stations. Furthermore, Hurricane Alex made landfall prior to the primary rainy season occurring in August and September in northeastern México (Shreve 1944). The low number of rain gauges (11 from daily CONAGUA sites, and 8 from automated stations) in the SCW limited the ability to characterize the spatiotemporal distribution of meteorological conditions during Hurricane Alex. For this purpose, the authors obtained meteorological fields from the North American land data assimilation system (NLDAS, Mitchell et al. 2004), a reanalysis product at 12 km, hourly resolution consisting of rainfall, air temperature, wind speed, solar radiation, pressure, and relative humidity. Following Robles-Morua et al. (2012, 2015), the authors applied a daily, mean-field bias correction to the NLDAS rainfall field over the period June 1 to July 31, 2010 for forcing both models in a similar fashion. Fig. 3(a) compares the daily total precipitation averaged over the SCW from rain gauges and NLDAS, indicating a general underestimation in the NLDAS fields. A mean-field, daily bias correction (Cázares-Rodríguez 2016) avoided the sharp discontinuities in the rain gauge patterns [Fig. 3(b)] and resulted in a large improvement in precipitation magnitude in NLDAS [Fig. 3(c)]. Nevertheless, inconsistencies in the subdaily timing of precipitation and other meteorological variables remain within the NLDAS fields used as forcing to the two hydrologic models.

Numerical experiments using the two hydrologic models were conducted over the period of June 28 to July 7, 2010. Rainfall in the SCW began on June 28, with the primary storm accumulations from June 30 to July 2. The flood peak at the basin outlet (Cadereyta Station) was estimated at nearly $\mathrm{4,300}{\mathrm{m}}^3/\mathrm{s}$ on July 2, with an estimated return period of 200–500 years based on historical analyses (Ramírez 2010). Fig. 5 compares the simulated streamflow at the hourly resolution used in both models at the basin outlet and Rompepicos Dam to the available observations. Because of uncertainty in the subdaily precipitation timing, cumulative discharge (${10}^6{\mathrm{m}}^3$) at the Cadereyta Station is shown in all cases [Fig. 5(a)] [see Cázares-Rodríguez (2016) for actual discharge]. Note the close match to the observed streamflow volume (i.e., cumulative discharge of $\sim 700\times {10}^6{\mathrm{m}}^3$) by tRIBS and the underestimation by HEC-HMS, as quantified by the bias ($B$) in Table 4. While both models have earlier flood peaks than the observations, the simulations exhibit high correlation coefficients (CC) of 0.70 (HEC-HMS) and 0.84 (tRIBS) with the streamflow record and relatively small peak errors (Table 4) amounting to 0.83% (HEC-HMS) and 5.84% (tRIBS) of the peak discharge. In addition, the Nash-Sutcliffe model efficiency (NSE) of 0.66 (tRIBS) and 0.38 (HEC-HMS) indicate a better model performance for the fully-distributed approach. Overall, the authors consider that the calibration exercise applied to each model yielded a good agreement to the CONAGUA records at the basin outlet when considering: (1) losses of streamflow data and their uncertain values during a destructive flood event, (2) differences in precipitation timing among rain gauges and NLDAS, and (3) similarities achieved in both model responses given their underlying differences. Having built confidence in the calibration at the basin outlet, testing at Rompepicos Dam shows that the models matched well the estimated peak outflow of $650{\mathrm{m}}^3/\mathrm{s}$ (dashed horizontal line) and have a similar peak inflow to the reservoir of $\mathrm{1,500}{\mathrm{m}}^3/\mathrm{s}$, thus achieving a peak reduction of 43% by the dam in both models. Unfortunately, there were no streamflow observation into the Rompepicos Dam for model evaluation purposes. Note that semidistributed and fully-distributed hydrologic model evaluations are not commonly conducted at internal locations in a basin (see Ivanov et al. 2004b). In addition, the close correspondence in the reservoir inflow and outflow behavior in the two models indicates the adequacy of the level-pool routing method developed for this study. Nevertheless, a few differences are noted between the models, namely an earlier rising limb and quicker recession limb in HEC-HMS, as explored subsequently.

The internal variability of the flood response in the two hydrologic models was compared by delineating three units (labeled Units 1, 2, and 3) and selecting a representative channel site (labeled Site I, II, and III) in each area. Fig. 6 presents the spatial location of the units and sites, while Table 5 summarizes the topographic, land cover, and infrastructure properties for each unit (also see the unit boundaries in Figs. 2 and 3). This delineation afforded the ability to separate regions in the SCW according to their contributions to the Hurricane Alex flood event: (1) Unit 1 is upstream of Rompepicos Dam and consists of high forest cover and complex terrain, (2) Unit 2 is a region with uncontrolled flow between Rompepicos Dam and the Entry to the City and is characterized by intermediate forest cover and terrain ruggedness, and (3) Unit 3 contains flatter areas with high amounts of impervious areas located downstream of the Entry to the City. The landscape differences between the three units help to explain the variations in the flood response between the hydrologic models at the three sites as quantified in Table 6. Note that Site I (Unit 1) exhibits the same pattern identified at the inflow of Rompepicos Dam, where HEC-HMS has a quicker response characterized by a lower time lag (10.9 h) as compared to tRIBS (17.4 h). In addition, Site I shows a large difference in the peak discharge between the two models ($729{\mathrm{m}}^3/\mathrm{s}$ in HEC-HMS and $533{\mathrm{m}}^3/\mathrm{s}$ in tRIBS). At Sites II and III, differences between HEC-HMS and tRIBS are progressively reduced, with nearly indistinguishable flood metrics for Site III (Table 6). This is explained by the different representations of slope (Fig. 4) in the two models, with the milder slopes in Unit 3 leading to smaller differences than the more variable slope conditions in Unit 1. As a result, the simulated flood response is more sensitive to the selection of a semidistributed or fully-distributed approach in areas where a greater aggregation of terrain properties is performed.

Larger differences between HEC-HMS and tRIBS in areas of complex terrain (Unit 1) as compared to flatter regions (Unit 3) were consistently found at other internal sites. To evaluate if other factors, such as impervious urban cover, might cause differences between the models, the authors evaluated the flood response along a set of main channel sites. Fig. 7 shows the locations and simulated flood events at four sites (Sites A, B, C, and D), while Table 6 provides a set of flood metrics for each model at these sites and the basin outlet. The HEC-HMS and tRIBS models show consistent differences along the main channel, with a noticeable decrease in the differences in peak timing among sites and an increase in the difference in peak discharge as the contributing area grows (i.e., from Sites A to D). An important change in the flood response occurs between Sites C and D located upstream and downstream of the main metropolitan area of Monterrey. The substantial growth in streamflow volume is attributed to runoff produced in the impervious regions in each model. Despite this contribution, the overall differences in flood peak amount and timing between the two models only varies moderately (Table 6). This suggests that the amount of urban cover does not explain the flood response differences between the semidistributed and fully-distributed models such that there is more sensitivity to the aggregation of terrain properties (slope) as compared to land cover conditions (impervious cover). Fig. 7 also demonstrates a few important features of the flood response represented in both models, namely (1) the contribution of Unit 1 (Site A) to the basin outlet is muted by Rompepicos Dam; (2) Unit 2 (Site B) has an important amount of uncontrolled flows (49 and 78% of the reservoir outflow volume in HEC-HMS and tRIBS); and (3) the urban area (between Site C and D) has a smaller contribution to the flood volume downstream of the city (Site D) as compared to the mountainous basin (only 42% of the total flood volume at the Cadereyta Station is produced in the urban area in both models). This latter finding is significant because stakeholder perceptions varied widely as to the relative importance of the mountainous basin and metropolitan area in generating the flood. In addition, both models consistently indicated the importance of uncontrolled flows from natural and urban tributaries downstream of Rompepicos Dam to the overall flood magnitude at the basin outlet.

Next, the authors compared the spatial variability of the hydrologic response in each model to identify the areas in the SCW with the highest contributions to the flood event. It is well known that interactions between terrain, soil, and vegetation patterns with meteorological forcing create preferential areas of soil moisture accumulation and runoff production (e.g., Smith and Hebbert 1979; Sivapalan and Wood 1986; Ivanov et al. 2004b; Mascaro et al. 2015). Fig. 8 presents a comparison of the time-averaged soil moisture conditions during the entire simulation period (June 28 to July 7, 2010) for the HEC-HMS and tRIBS models. Because the variables of interest are different for each formulation (i.e., saturation fraction in HEC-HMS, and root-zone relative moisture in tRIBS) the values shown in the colorbars do not match. Despite this, qualitative comparisons can be made among the semidistributed and fully-distributed models and the spatial patterns within each model can be assessed relative to the landscape properties and meteorological data. The primary trends captured in both models are (1) relatively drier soils in areas upstream of Rompepicos Dam (Unit 1) due to lower rainfall accumulations [Fig. 3(c)]; (2) a large contrast between impervious urban areas and surrounding rural lands exhibiting higher wetness [Fig. 2(c)]; and (3) relatively wetter soils in valley bottoms consisting of shrublands and secondary shrublands. Notably, the representation of soil moisture is more highly resolved in tRIBS, where spatial differences between drier mountain ridges and wetter valley bottoms are depicted well.

Fig. 9 presents a comparison of the total runoff (mm) produced during the simulation by each model and the runoff associated with two of the primary mechanisms in tRIBS (infiltration-excess and saturation-excess runoff) (Ivanov et al. 2004a; Vivoni et al. 2007). Overall, a similar runoff pattern was produced in the semidistributed and fully-distributed approaches, driven in large part by the rainfall accumulation [Fig. 3(c)]. Note that the NLDAS 12 km pixels are visible in the runoff distribution from tRIBS as the computational elements are much smaller than the meteorological forcing, as opposed to the rainfall aggregation occurring in HEC-HMS. Nevertheless, the close correspondence between the approaches builds confidence in the runoff capabilities of the fully-distributed model with respect to the modeling tool commonly accepted by local stakeholders (Ramírez 2010). In addition, tRIBS adds a considerable amount of spatial detail that captures the high-resolution interaction of rainfall characteristics with soil and terrain properties in the SCW, which ultimately leads to runoff production in a landscape and discharge in the stream network. Aggregation of the watershed soil and terrain information in HEC-HMS results in a substantial loss of information. Furthermore, the fully-distributed model is able to simulate spatial patterns of different runoff mechanisms indicating that (1) the metropolitan area of Monterrey produces runoff via the infiltration-excess mechanism due to the large amounts of impervious surfaces, (2) drier areas in the upper basin [Fig. 8(b)] have moderate amounts of infiltration-excess runoff due to the formation of shallow layer of surface saturation in regions of moderate to high slope (also see Ivanov et al. 2004a), and (3) wetter areas in valley bottoms and mountain ridges [Fig. 8(b)] that reach saturation of the entire soil profile primarily produce saturation-excess runoff. The ability to distinguish between these runoff mechanisms can help select among different flood mitigation strategies and identify key locations for the placement of hydraulic infrastructure options.

The authors assessed the impact of the current hydraulic infrastructure in the SCW and the potential for flood mitigation afforded by the proposed scenarios. While similar efforts have been conducted with HEC-HMS (e.g., Emerson et al. 2003; Robles-Morua et al. 2015), this study reports the first use of reservoir routing and infrastructure scenarios with tRIBS. Fig. 10 shows the locations and upstream contributing areas (CA) of the hydraulic infrastructure scenarios, with additional details on the site locations presented in Cázares-Rodríguez (2016). The large dam at the Entry to the City has a high upstream CA ($\mathrm{1,129}{\mathrm{km}}^2$ or 62% of the SCW area) and serves as a clear demarcation between the upper rural area (Unit 1 upstream of Rompepicos Dam and Unit 2) and the lower urban (Unit 3) area of the watershed. The location of the large dam (also labeled as Site B in Fig. 7 and Table 6) coincides with a switch between wet valley areas producing saturation-excess runoff and downstream urban areas where infiltration-excess runoff is the dominant mechanism. In contrast, the three small detention dams have much lower CA ranging from 106 to $170{\mathrm{km}}^2$, placed at strategic locations in terms of uncontrolled runoff production during Hurricane Alex as simulated by the two hydrologic models. Two of the three detention dams correspond to analysis locations (Site II and III) in Fig. 6 and Table 6. Both stakeholder strategies (large dam at the Entry to the City dam and three small detention dams) retained the existing Rompepicos Dam within the two models.

Fig. 11 summarizes the effects of the hydraulic infrastructure scenarios from both models at three watershed locations of increasing CA that are relevant to the Monterrey metropolitan area: (1) Entry to the City, (2) Site C, a confluence upstream of the main urban areas (Fig. 7), and (3) Cadereyta Station at the basin outlet. The scenario labeled Current Conditions refers to the calibrated flood responses from the semidistributed and fully-distributed approaches that contain the flood mitigation achieved by the Rompepicos Dam. The favorable effects of the existing infrastructure are captured in the scenario labeled No Dams, where Rompepicos Dam results in a reduction of the peak discharge of approximately 39% at the Entry to the City and 14% at Cadereyta Station, an indication that the Hurricane Alex flood event would have been more severe (i.e., exceeding $\mathrm{5,000}{\mathrm{m}}^3/\mathrm{s}$ at the basin outlet) without the existing hydraulic infrastructure. The two stakeholder-driven mitigation strategies varied in their effectiveness upstream of the Monterrey metropolitan area (Entry to the City and Site C), whereas differences were reduced at the Cadereyta Station because of the contributions from impervious urban areas downstream of the hydraulic infrastructure. The inclusion of a large dam at the Entry to the City had the effect of decreasing and delaying the flood peak at all downstream locations in both models because of its capacity to separate in time the contributions from upper areas (Units 1 and 2) and lower areas (Unit 3) to the overall flood wave. At the Cadereyta Station, a reduction of the peak discharge of 17% (tRIBS) and 23% (HEC-HMS) was obtained due to the proposed dam at the Entry to the City. Note that Hurricane Alex would have led to a water level exceeding the main Creager spillway at the Entry to the City site (see sharp discharge rise in Large Dam scenario on July 2) if similar characteristics to the Rompepicos Dam were used in the design. As a comparison, inclusion of the small detention dams had a smaller mitigation effect at the three downstream locations, with a lower peak reduction at the Cadereyta Station (12% in tRIBS and 10% in HEC-HMS) and a lower overall effect on the flood peak timing. At Site C where the effects of all three structures can be determined, the flood peak discharge was reduced by 17% (tRIBS) and 20% (HEC-HMS) relative to the current conditions. Thus, the strategic placement of smaller hydraulic structures in regions with uncontrolled flows, as illustrated through the example of the three small detention dams (see additional details in Cázares-Rodríguez 2016), holds promise in collectively achieving an important degree of flood mitigation upstream of Monterrey as shown through this model comparison.