Street-Scale Modeling of Storm Surge Inundation along the New Jersey Hudson River Waterfront (2024)

Keywords: Flood events; Hurricanes/typhoons; Storm surges; Forecast verification/skill; Numerical weather prediction/forecasting; Ocean models

1. Introduction

Storm surges are among the world’s most costly and deadly disasters, and recent hurricanes like Sandy and Katrina and Typhoon Haiyan highlight the threat worldwide. Modeling inundation in coastal cities and towns (defined as the area within 100 km of a coastline) has become important because the world’s inland rural population is moving to the coast (Creel 2003). Over 39% of the U.S. population lived in coastal shoreline counties in 2010 (NOAA 2013). Over 50% of the world population lives in coastal areas, and this percentage is projected to keep increasing for the foreseeable future (Creel 2003; Tibbetts 2002). In the largest coastal cities, the 136 port cities around the world that have more than 1 million inhabitants, there is a population of 400 million people (Hallegatte et al. 2013).

Increasing damage from coastal flooding is one of the most certain impacts of climate change, with storm surges coming on top of rising sea levels, and with the potential for intensified storms and increased rainfall in the northeastern United States (Walsh et al. 2014). Sea level rise is expected to accelerate over the twenty-first century, primarily due to increasing expansion of warming seawater and accelerated melting of land-based ice sheets. A conservative estimate of 30–60 cm for New York City, New York (NYC), by 2080 will change a 100-yr flood event to a 30-yr flood event; the latest localized projections show a 25% chance of sea level rising more than a meter over this period (Horton et al. 2015). Using recovered archival tide gauge data back to 1844 for New York Harbor, Talke et al. (2014) showed that flood levels in New York Harbor have been increasing due to rising sea levels and also due to increasing storm tides, the latter for unknown reasons. The annual likelihood of overtopping seawalls has increased from below 1% in the mid-1800s to about 20%–25% today.

On 29 October 2012, Hurricane Sandy overwhelmed the most densely populated region of the United States. The resulting damage was widely experienced within the tri-state area. During Hurricane Sandy’s pass through the New Jersey Hudson River Waterfront cities of Hoboken, Jersey City, Weehawken, and Bayonne (Hudson waterfront) severe impacts occurred. The four cities are among the most densely populated cities in the United States. They lie across the Hudson River from Lower Manhattan and are bordered by water on two sides—with the Hudson River and Upper New York Bay on the east, and the Hackensack River and Newark Bay on the west (Fig. 1). In Jersey City, for example, about 75% of the population lost power because of Sandy storm surge flooding, with many residents not having gas and electricity restored for more than a week. Some 2500 residents sought shelter due to a lack of power, water, and heat. With 50 000 people living in one square mile, Hoboken is the fourth most densely populated municipality in the United States. Many of its residents were without power for nearly 2 weeks after the storm. Sandy crippled the Port Authority Trans-Hudson (PATH) line, a 24-h subway that last year ferried 76.6 million passengers between Manhattan and New Jersey. The entire system was out for 2 weeks. A link to the World Trade Center was out for 4 weeks, and the Hoboken line restored service months later. All repairs and projected costs to the PATH system are expected to ultimately exceed $700 million. The costs associated with Hurricane Sandy in Jersey City alone could easily approach $100 million, and the cost associated with damages to city-owned property and equipment alone is estimated at approximately $23 million.

The Hudson waterfront cities have much of their land within the new draft 1%-flooding-probability-per-year (“100 years”) flood zone (FEMA 2014), yet many of the buildings are high-rises, with steel beams into the bedrock or attached row houses and brownstones that generally cannot be raised. As a result, the Hudson waterfront is an excellent example of an urban coastal region that is badly threatened by sea level rise and storm surges. To address possible inundation from storms of the future, a modeling methodology is developed herein and an analysis is carried out with the overriding goal of predicting the inundation (herein defined as land that usually has no water over it) likely to occur from a storm surge event. An accurate model will improve the capacity of Hoboken, Jersey City, Weehawken, and Bayonne to adapt to coastal flooding both from storms and a rising sea level. It is important that we consider these four adjoining cities together in the analysis because it is likely that a protective measure in one city may adversely affect its neighboring city. The methodology developed herein is certain to be useful in addressing flooding in coastal urban cities worldwide.

Storm surge modeling—that is, the modeling of the sea level anomaly caused by winds and spatial variations in atmospheric pressure—has had a long history beginning with the pioneering work of Jelesnianski (1965) and Heaps (1983). The early models were two-dimensional, vertically integrated, often linear, parameterizing key processes into parameters that were tuned to give good agreements between the observations and the model simulations. These models have evolved to include unstructured grids that resolve to a high degree a region’s topography and bathymetry and are able to incorporate basin-scale domains that simplify specification of the open ocean boundary conditions (Westerink et al. 2008; Resio and Westerink 2008). Today, models that simulate storm surges are increasingly three-dimensional because of their ability to provide information on the vertical structure of storm-induced currents. Near the seabed, this allows a more physically based specification of bottom friction than can be obtained from a two-dimensional parameterization, based on depth-averaged currents (Zheng et al. 2013).

The availability of high-resolution topographic maps of the elevation of land above common datum digital elevation models brought the idea of inundation modeling. There are two basic methods for addressing inundation from storm surges just beyond the land water interface. The simplest flood model in use today is static mapping—a method in which flood waters are assumed to spread out horizontally to cover all land areas of equal or lower elevation. Higher sea level elevations result in greater floodplain areas, with the extent of landward flooding dependent on elevation and slope of land, presence of manmade structures, permeability of soils, vegetation, and other impediments to movement of water.

A more advanced version of static mapping considers hydraulic connectivity (NOAA Coastal Services Center 2012), removing unrealistic pools in the inner low land areas caused by blindly comparing the heights between the ground and the extended water surface. Hydrodynamic inundation models, the second of the two methods for simulating inundation, although more complex and more computationally demanding, are able to capture the flooding/drying physics and yield more meaningful results.

A potential shortcoming of the static approach is that the horizontal movement of flood waters over land areas is determined by friction, wind, and other dynamic factors that are likely captured better using hydrodynamic modeling. This is the reason why the Federal Emergency Management Agency (FEMA) utilizes hydrodynamic models for its flood mapping studies (FEMA 2014) and why NOAA uses them for forecasting neighborhood flooding during hurricanes. Moreover, prior hydrodynamic modeling studies of New York Harbor have shown a good capability for models to reproduce past storm tide events with typical accuracy within 0.15 m. Compared to results of the dynamic modeling approach, the static approach to projecting coastal flooding from hurricanes underestimates flood heights and flood zone areas for several locations in the New York metropolitan region. For example, the dynamic modeling approach shows Midland Beach in Staten Island to be more vulnerable than the static model depicts. (Orton et al. 2015).

The extension of storm surge models to inundation models where land that was once dry becomes flooded has come about because of better algorithms used for the treatment of the flooding/drying process. The models typically handle the flooding and drying process in a phenomenological fashion requiring wet–dry checks and minimums in the depth allowed (Medeiros and Hagen 2013). In this paper, we demonstrate a flooding/drying algorithm that is based on the Princeton Ocean Model (POM) of Blumberg and Mellor (1987) as modified by Oey (2006). The algorithm dynamically determines whether a grid is wet or dry and accordingly includes or removes it into or from computation. It balances the trade-offs among computational costs, physical authenticity, and realistic results.

2. New Jersey Waterfront inundation modeling system

The waterfront model is a Stevens Institute Estuarine and Coastal Ocean Hydrodynamic Model (sECOM; Blumberg et al. 1999; Georgas and Blumberg 2010) application, nested into the larger New York Bight sECOM [New York Harbor Observing and Prediction System (NYHOPS); Fig. 2). sECOM is a three-dimensional, free surface, hydrostatic, primitive equation estuarine and coastal ocean circulation model. Prognostic variables include water level, 3D circulation fields (currents, temperature, salinity, density, viscosity, and diffusivity), significant wave height, and period. It is the successor model to the Estuarine Coastal Ocean Model–Princeton Ocean Model (ECOM–POM) combination that is in use by almost 3000 research groups around the world with over 1000 papers having been published with them as the modeling engine (Blumberg and Mellor 1987). The sECOM operational forecast application to the New York–New Jersey harbor estuary and surrounding waters (NYHOPS) is found online (http://www.stevens.edu/maritimeforecast) and dates back to 2006 (Bruno et al. 2006; Fan et al. 2006; Georgas et al. 2009a; Georgas 2010), and includes forecasts of chromophoric dissolved organic matter and associated aquatic optical properties through coupling to a row column assessing the effects of submesoscale ocean parameterizations (Aesop; RCA)-based water quality model (Georgas et al. 2009b).

In its NYHOPS application to the waters of New York and New Jersey (Georgas et al. 2009a; Georgas and Blumberg 2010; Georgas 2010), the computational domain is discretized on an Arakawa “C” finite-difference curvilinear grid (147 × 452 horizontal cells, 15 068 of which are designated as water). The NYHOPS grid (Fig. 2) encompasses the entire Hudson–Raritan (New York–New Jersey Harbor) estuary, the Long Island Sound, and the New Jersey and Long Island coastal ocean. The resolution of the grid ranges from approximately 7.5 km at the open ocean boundary to less than 50 m in several parts of the NY–NJ harbor estuary. In the vertical, the model uses a sigma-coordinate system with bathymetrically stretched sigma layers to permit better representation of bottom topography. The current vertical resolution of the NYHOPS grid is 10 sigma (bottom following) layers at depths shallower than 200 m, providing forecasts at 150 680 points averaged every 10 min. Surface gravity waves are modeled in NYHOPS using a parametric model of wind-wave growth, propagation, and frictional decay (Donelan 1977). The bottom stresses are computed using the Grant and Madsen (1979) wave-current boundary layer approach (1977). Explicit wave roughness is included via the surface wind stress being computed with the Taylor and Yelland (2001) drag formulation. Several comprehensive skill assessment studies have been carried out (Fan et al. 2006; Georgas et al. 2007; Bhushan et al. 2010; Georgas and Blumberg 2008, 2010; DiLiberto et al. 2011) and in each case NYHOPS performance has been quite good.

The New Jersey waterfront inundation model has a constant 3.1-m horizontal resolution. It is quite sufficient to resolve the main avenues in the region, which are typically 20–25-m wide with 130-m-long blocks. The cross streets are 14 and about 60 m long. The grid and bathymetry are shown in Fig. 3. The flooding and drying model’s external time step is 0.1 s in its 2D barotropic mode. The bottom drag coefficient for the land areas is taken as 0.025. This is a factor of 10 higher than values used in the offshore waters and as such represents the higher friction from the roads’ rough surfaces, curbs, railroad tracks, and street side vegetation, relative to a smoother marine or estuarine sediment substrate.

The collection and application of bathymetric and topographic data on such small scales were quite involved. A 10-ft (3 m)-resolution digital topobathymetric DEM with a vertical accuracy of 0.185 m that used lidar data as its basis (FEMA 2014) was used as a base map. For the two major urban areas, Hoboken and downtown Jersey City, the building blocks and the places that are deemed resistant to flooding were located based on aerial images from Google Earth and set to infinitely high. Local corrections were included for the piers and for the important area around the New Jersey Transit terminal in Hoboken from ground surveying. The relative elevation of the several piers in the region was estimated by measurement tape and the pier elevation was then adjusted according the adjoining street elevations found in the digital terrain model (DTM). The NJ Transit area was handled similarly: the heights of the walls and doorways were measured relative to the ground and then adjusted again according to the adjoining DTM street elevations. The bathymetric data of the very nearshore waters were determined using the personal watercraft–based Stevens Dynamic Underwater and Coastal Kinematic Surveying System (DUCKS) (Miller et al. 2009). All the bathymetric and topographic data were assembled, quality controlled for consistency by visual inspection, converted to the North American Vertical Datum of 1988 (NAVD88) and then placed on a 3.1-m square grid for use with the modeling. The bathymetric and topographic data are shown in Fig. 3.

Wind and pressure fields representing Hurricane Sandy were obtained from the Rutgers, The State University of New Jersey (RU) WRF (RU-WRF) Model. Their highly refined mesoscale hindcast methodology provides wind, pressure, and surface heat flux variables on a 3-km resolution and hourly in time. Interior streamflow data were from USGS. The resulting time series of NYHOPS water level used to force the offshore boundary of the fine-resolution waterfront model nest is shown in Fig. 4. The NYHOPS total water level results forced by the RU-WRF winds were compared with observations at stations in and around the New York metropolitan area. The NYHOPS predictions were excellent with root-mean-square error between model and observations of 0.18 m in upper New York Harbor and the lower Hudson River, an error similar to the error in the base DTM (Georgas et al. 2014). The model simulation begins at 0000 eastern daylight time (EDT) 28 October and ends at 2345 EDT 30 October.

3. Model validation

Multiple types of data sources were used to validate the model. By far the greatest set of “data” came from a crowdsourcing initiative in Hoboken. E-mail announcements asking for photographs, videos, or just recollections of flooding were sent out to thousands of people. Water level information sought were location of the site observed, height of the water, and the time it was observed. Photographs were found to serve the model validation best. Hundreds of photographs were received. Figure 5 was typical of a verifiable crowdsourced photograph; determining the water level, location, and time was easy from this photograph of the New Jersey Transit terminal clock tower. Flood heights were determined from the useful photographs by going to the site and measuring the water elevation relative to doors, walls, and vehicles. Unfortunately, most survey responses were of little use. They typically lacked an accurate time registration and often a precise location. People with very useful photographs could not figure out how to off-load them from their cell phones. Phone photograph files are apparently private; the people did not want us to help them with the off-load. Many witnesses were interviewed; most had great stories of the flooding but little specific information that could be used for model validation.

In total, 56 watermarks were used; there were 19 USGS-verified high watermarks (McCallum et al. 2012; USGS 2014), 26 verbal story watermarks and 11 watermarks from photographs or videos (see Fig. 6 for the locations of all the watermarks). Only the USGS watermarks measure the highest water level. The comparison of the model results to the watermark observations is shown in Fig. 7. The error bars (vertical lines in Fig. 7) represent the estimated errors in the height of water, whereas the horizontal error bars depict possible errors in the time of the occurrence of the watermark itself. For example, if a witness estimated that a water level occurred at around 2000 EDT or more specifically at a time between 1945 and 2015 EDT, the two ends of the horizontal error bar are the maximum and the minimum water levels reached during this estimated time range in the model.

The correlation coefficient (R²) between the watermark observations and the model results is 0.93 and the average error is 0.05 m. Accounting for the uncertainty in the observations, the standard deviation of the residual error is 0.07 m. The simulated inundation depths at 78% of the data measurement locations have less than 20% error. Inundation depths in excess of 2 m were predicted quite well. In general, the lower watermarks have a larger time error, while the higher high watermarks have a greater elevation error estimate. The greatest differences between the observed inundation and the modeled inundation occur at locations very close to the Hudson River, where the water level changed very quickly during Hurricane Sandy. Accurate time estimates from the crowdsourced information were difficult to determine at these locations.

The limited crowdsourced observations in Hoboken also provided 16 flood edgemarks (Fig. 6), locations at which the floodwaters stopped. They suggest that the model was able to reproduce the flood extent to within 20 m.

To illustrate that the model was able to capture the spatial and temporal variation of the storm surge in Hoboken, the time histories of water level both modeled and observed at all the stations farthest from the waterfront are plotted in Fig. 8. Being farthest away from the waterfront provides the greatest challenge to the model, as the path from the waterfront to the particular station is often convoluted. The station in the southeastern corner of Hoboken, while close to the waterfront, is included, as it has the highest water modeled. The time series plots show good agreement. The peak surge, an aspect hard to obtain from observations alone, is quite apparent.

4. Flood pathways

The maximum water levels that were reached along the Hudson River Waterfront during Hurricane Sandy are shown in Fig. 9. The flooding was widespread and quite deep. An analysis of the flood pathways based on this flooding is useful for resiliency planning. Figure 9 also presents the flood pathways into Hoboken and Jersey City as derived from the model simulation. The flood started to enter the back of Jersey City from the western end of Morris Canal about 5 h before the Hudson River water level peaked, and it started to enter the back of Hoboken from the Long Slip Canal and the Weehawken Cove about 3 h prior to when the river water level was at its highest. In the back of the cities, the maximum water level over ground was reached shortly after the peak of the river, but the maximum extent of the flooded area was not reached until about 1 h after the peak.

In total, approximately 1 764 000 m³ of floodwater entered into Hoboken. Half of that volume entered Hoboken south of the Hoboken train terminal. Along this path, a sixth of the water volume, or 295 000 m³, did not make it across the terminal into the streets of Hoboken. A portion of the floodwater remained in the terminal or flowed south into Jersey City. The remaining 583 000 m³ of floodwater flowed either across the train terminal or south out of Long Slip Canal into Hoboken. The two major entry points were the open space west of the New Jersey Transit building (accessing Observer Highway between Willow and Park Avenues) and the northern end of Marin Boulevard—371 000 m³ of water entered through the former point and 212 000 m³ through the latter point.

At the north entry point, 723 000 m³ of water flowed into Hoboken from Weehawken Cove. Two-thirds of this volume of water entered across the northwest bank of the cove and the remainder across the southwest bank. These two volumes of water merged in the northwest portion of Hoboken between Fifteenth and Sixteenth Streets and propagated south.

Floodwaters from the north and south sides of Hoboken met near Seventh Street; however, a net flux of more than 87 000 m³ of water pushed south toward Observer Highway. Small volumes of floodwater entered Hoboken through the Erie-Lackawanna Park at the south end of Hoboken and at the eastern end of Fifteenth Street; the floodwaters there were constrained by the higher topography along the east side of Hoboken and contributed little to the flooding in the interior of the city. It is important to note that there are flood pathways between Hoboken and Jersey City. Clearly evident in Fig. 9, there is flow to the south, out of Hoboken and toward Jersey City.

The flood pathways into Jersey City are also shown in Fig. 9. Like Hoboken, floodwaters entered the city at low points located at the north (Long Slip Canal) and south (Morris Canal) boundaries. Unlike Hoboken, floodwater also entered into Jersey City at low points along the Hudson River. Flooding from Morris Canal started approximately 3 h before the peak of the surge, when water levels begin to exceed 1.8 m above NAVD88. Water flowed from Morris Canal northwest toward the New Jersey Turnpike and Grand Avenue. Approximately 2.5 h before the surge peak, floodwater began to enter Jersey City from the Hudson River at Exchange Place and water began to flow into Liberty State Park from Morris Canal.

An hour and a half before the surge peak, water began to enter the northern part of Jersey City from the Long Slip Canal along Marin Boulevard and from the Hudson River at Newport Yacht Club and Marina. Floodwaters from Exchange Place flowed west down Columbus Drive and water began to enter the city from Liberty Harbor Marina located on the north side of Morris Canal. At this time, most of the northern section of Liberty State Park was underwater and water depths were up to 1 m west of Morris Canal.

At the peak of the surge, the floodwater entering the city from Liberty Harbor Marina and the west end of Morris Canal merged with the floodwater flowing from Exchange Place and Newport Yacht Club and Marina, inundating the eastern side of Jersey City under up to 1 m of water. Northwest of Morris Canal water depths reached 1.8 m and water began to extend west toward Communipaw and north toward Newark Avenue, flooding the historic downtown. Floodwater that entered from the Long Slip Canal flowed west along Eighteenth Street, inundating the northwest portion of Jersey City. After the peak surge, floodwaters continued to flow north from the historic downtown area until they merged with the floodwaters in the northwest portion of the city.

The flood impacts are quite dependent on location as shown in Fig. 8. What becomes apparent from an analysis of the model results is the entire Sandy event lasts about 4 h, from 1900 EDT 29 October 2012 to about 2300 EDT that same day. Low-lying areas tended to trap the floodwaters, unfortunately, and slowly drained through a clogged sewer system over a period of days in addition to being pumped out. The model validation ends soon after the peak has occurred because the model does not include ground infiltration, sewers, or pumps.

5. Discussion

The comparison of the model results to the observations shows excellent agreement. This is due to the robust flooding and drying physics of sECOM, the high-resolution and accurate digital terrain data assembled for this study, the fine resolution used in the model, and the high-fidelity forcing functions brought to this study. The close agreement found in this paper provides a high confidence in the use of the model for overland inundation prediction.

Our flood edgemark validation appears similar to the results of Wang et al. (2014), in a study of Sandy inundation in New York City, who found a 30-m mean absolute difference of the maximum extent of inundation between modeled and the data-derived edgemarks. Their study was a hybrid simulation in that the model grid consists of a 200 m × 200 m resolution square base grid with an embedded 5 m × 5 m resolution subgrid. Bathymetry was from NOAA surveys and coastal relief models, and the land topography was from USGS lidar surveys and the Open NYC Building Inventory. Observations at three locations showed model results that were consistent with that data. However, they report that USGA data (McCallum et al. 2012) was used for validating the spatial extent of the inundation. However, the dataset they refer to was actually created by FEMA Modeling Task Force (2014) and is mischaracterized in the Wang paper as spatial flood observations—it is actually an estimate of spatial flood boundaries made by extrapolating (bathtubbing) elevation observations from discrete USGS high watermarks, in many cases over distances of over 1 km. It is not possible to know the actual accuracy of their flood area model results unless they compare it to the actual high watermark point data.

A drawback to the use of this new inundation model is with computational speed. The runs in this paper took about19 h to complete using 16 processors on a high-end Dell server. Running with a limited number of processors is insufficient for taking full advantage of the new computational capabilities that a large, high-resolution sECOM coastal ocean domain can deliver. Now efforts are underway following Jordi and Wang (2012) to parallelize the code. Two-dimensional data decomposition of the horizontal domain is used with a halo of ghost cells to minimize communication between processors. Communication consists of the exchange of information between neighbor processors based on the message passing interface (MPI) standard interface. The netCDF-4 library is also being implemented with parallel features to achieve a high efficient input and output (I/O).

6. Conclusions

A new, high-resolution, hydrodynamic model for the Hudson River Waterfront on the New Jersey side facing Manhattan including four municipalities—Weehawken, Hoboken, Jersey City, and Bayonne—has been developed and validated. The New Jersey Waterfront Inundation Model has a constant 3.1-m resolution and is nested within the three-dimensional NYHOPS model at its offshore open boundary, influenced by estuarial tide and storm surge. Flooding and drying of land features in the model’s external time step is as low as 0.1 s in its 2D barotropic mode. This mode provides for the dynamic prediction of depth-integrated flood elevations and velocities across land features during inundation events. The model was calibrated using the NYHOPS hindcast of Hurricane Sandy. The hindcast utilized Sandy over ocean wind field and atmospheric pressure data, offshore wave and tidal boundary forcing, atmospheric heat fluxes, and interior streamflow data.

Watermarks at 56 locations and 16 edgemarks from a combination of USGS data archives and crowdsourced photographs, videos, and stories were obtained. A comparison shows that the model is capable of hindcasting overland water elevation accurately. The correlation coefficient (R²) between the watermark observations and the model results is 0.93. The standard deviation of the residual error is 0.07 m. The simulated inundation levels at 78% of the data measurement locations have less than 20% error. Inundations in excess of 2 m were predicted quite well. Because the model was able to capture the spatial and temporal variation of water levels in the region observed during Hurricane Sandy, it was used to identify the flood pathways and to suggest where flood-preventing interventions could be built.