Zhuo Liu, Ph.D., Flood Data Scientist at One Concern, Inc.

Note to reader: The model described below is in development by One Concern, as part of the company’s ongoing research into the challenging and increasingly important field of coastal flood modeling. All results are preliminary and for discussion purposes only.

In this article, we share the results from applying an advanced large-scale hydrodynamic model, SCHISM (Semi-implicit Cross-scale Hydroscience Integrated System Model) with hurricane forecasts as inputs (see left panel of GIF below), during Hurricane Dorian to predict the storm tide, a combination of astronomical tide and storm surge, along the coastlines of the Bahamas and entire U.S. East Coast (see the right panel of GIF below). The predicted potential impacted areas were compared with satellite images in Grand Bahama Island, and the predicted peak water levels were compared with NOAA tidal gauge observations along the U.S. Southeast Coast. The predicted coastal water level shown can be used to drive local inundation models to provide coastal flooding forecast.

The highlights of this case study are:

  • One Concern’s large-scale SCHISM model was used to predict storm tide, the total water levels resulting from the combination of storm surge and astronomical tide, during Hurricane Dorian. The large-scale modeling domain we used was able to cover Dorian’s potential impacts along the Bahamas coasts and the entire U.S. East Coast.
  • The model predicted a devastating storm surge along the coast of Grand Bahama Island. This prediction was made 2 days in advance of Dorian’s direct hit to the Bahamas, and subsequently was demonstrated to be consistent with the actual impacted areas captured by satellite images.
  • The model predicted the peak storm tides along FL, GA, SC, and NC coastlines 12–24 hours in advance, with average error < 0.3 m against NOAA preliminary observations.

Background

To set the stage, there are a few definitions and pointers about hurricane predictions we will want to clarify before sharing the results of our model.

Storm tide vs. storm surge

First, it is key to understand what we mean by storm tide and storm surge to interpret our results. Based on NOAA’s definition, storm surge is the abnormal rise in seawater level during a storm, measured as the height of the water above the normal predicted astronomical tide. Storm tide is the total observed seawater level during a storm, resulting from the combination of storm surge and the astronomical tide. In reality, storm surge only makes up a part of what causes water levels to rise along the coast during a hurricane. The storm surge with concurrent high tide is usually much more damaging than during low tide condition, so it is very important to model astronomical tide along with storm surge to make a realistic water level prediction during a storm. For example, if the storm surge is 1.5 m and predicted high tide is 1m above MSL, the resulting total water level would be 2.5 m above MSL, which means the coastal areas lower than 2.5 m above MSL would be under the potential risk of storm surge flooding. However, if the storm surge is happening during low tide (e.g. 1 m below MSL), the storm tide would be only 0.5 m above MSL.

Inherent uncertainty in hurricane forecasts

It is very important to note that our storm surge model doesn’t predict whether or when a hurricane will make landfall in a certain place, but rather predicts the potential storm tide given inputs from hurricane forecasts. The left panel of the GIF above shows a total of 10 selected NOAA’s forecast tracks published between August 28 and September 5 during Dorian. In the figure, it is apparent that the hurricane forecast track changed frequently and dramatically during this event. For example, in early August 31, the forecast completely changed from “Dorian making a landfall in FL” to “Dorian NOT making any landfall.” High uncertainty in hurricane forecasts means it is important to run predictions as soon as the forecast is updated to provide the most possibly accurate results.

Results

Due to high uncertainty in the input hurricane forecasts, we decided to monitor the storm and run 7-day storm tide predictions each day beginning August 28 to have the most up-to-date results. Thus, the predicted results for each seven day period would change daily, as hurricane forecasts were updated.

Each one of these 10 forecast tracks was generated by NOAA integrating various atmospheric forecast models including GFS (Global Forecast System). We chose GFS forecasts as SCHISM’s atmospheric inputs because GFS is being used as NOAA’s flagship weather model. Each of our SCHISM predictions were based on the GFS forecast used to generate the NOAA’s published forecast track. This can explain why for each of the 10 predictions made the storm surge prediction changed frequently during this event as the forecast track changed. Below, we presented all of 10 predictions to show you the dynamics of storm surge prediction during a hurricane event (shown in the right panel of the GIF above). The accuracy and confidence in the predictions are typically expected to improve as updated hurricane forecasts are used.

We built an interactive map (shown at the beginning of this post) to present the 10 predictions of peak water levels along Grand Bahama and U.S. coasts during Hurricane Dorian. The user can zoom and pan around to see the max water level estimates along with the categories and locations of Dorian which are marked along the preliminary best track published by NOAA’s National Hurricane Center.

  • Prediction #1 (08/28 1200UTC — 09/04 1200UTC) used a GFS forecast initialized at 08/28 1200 UTC (8AM ET). At this time, most hurricane forecast models predicted Dorian would make landfall in Central FL in 5 days on Labor Day, September 2. Based on geophysics and coastal oceanography, the larger impacts would be on the east side of the storm track between the coastlines of Savannah, GA and Jacksonville, FL.
  • Prediction #2 (08/29 1200UTC — 09/05 1200UTC) started on August 29 using GFS initialized at 1200UTC (8AM ET), and showed similarly higher water levels along the GA coast since the forecast track didn’t change too much from #1. At this time, Dorian was still a Category 1 Hurricane.
  • Prediction #3 (08/30 0600UTC — 09/06 0600UTC) started on August 30 used GFS initialized at 0600UTC (2AM ET). The potential impacted areas shifted to Grand Bahama Island as the forecast track indicated a potential landfall at the island. Dorian was a Category 2 Hurricane at this time.
  • Prediction #4 (08/30 1800UTC — 09/06 1800UTC) used GFS initialized at 1800UTC (2PM ET) on August 30 when Dorian quickly escalated to Category 3 Hurricane, and later most hurricane models agreed that Dorian would steer away to the Northeast and would not make any landfall on the FL coast. It would travel in parallel to the coastlines of FL, GA, and SC. This SCHISM prediction showed that an extremely high storm tide (> 3m) would occur along the coast of Grand Bahama Island (shown below). In fact, Dorian started to hit the Bahamas heavily on 1800UTC, September 1 as a Category 5 Hurricane and caused huge storm surge flooding in the island. Prediction #4 showed the potential impacted areas due to extremely high storm tide 2 days in advance of Dorian’s hit to the Grand Bahama. The prediction was consistent with later satellite images captured by ICEYE.
Figure 3: Raw output from our predicted maximum water level (m above MSL) near Grand Bahama Island (displayed in QGIS software).
Figure 4: Prediction #4’s outputs showing peak water level along the coastlines of the Grand Bahama Island and Great Abaco Island (east of Grand Bahama Island).
Figure 5: ICEYE’s satellite image of Grand Bahama Island during Hurricane Dorian on September 2. This shows that large sections of the island, including its main airport, are now flooded due to extremely high water level. The yellow lines mark the shorelines before Dorian’s storm surge flooded the island, and the white lines mark roads (Source: OpenStreetMap). We use this to compare against the predicted high water level areas above from our model. You can find other recreated satellite images from CNNNPR and the Washington Post.
  • Predictions #5 (08/31 1800UTC — 09/07 1800UTC) and Prediction #6 (09/01 1800UTC — 09/08 1800UTC) were consistent with Prediction #4 that Grand Bahama Island would experience huge impacts from Dorian in terms of dangerously high water level (>3 m above MSL). This prediction also indicated that FL, GA, SC, and NC coasts would also feel Dorian’s impact to different degrees.
  • Predictions #7 (09/02 1200UTC — 09/09 1200UTC) and Prediction #8 (09/03 1200UTC — 09/10 1200UTC) showed high water level would start to appear on the U.S. Southeast Coasts. As Dorian was moving slowly (< 10 mph) to the North, it was “parking” over Grand Bahama and continued damaging the Bahamas throughout September 3.
  • Prediction #9 (09/04 1200UTC — 09/11 1200UTC) used GFS forecast initialized on 1200UTC on September 4 and showed that Dorian would impact most of US Southeast Coast as the forecast track would be very close to the coastline. The model predicted 1.7 m peak water level along the coast of Jacksonville Beach, FL (vs. 1.4 m observed by NOAA), 1.7 m along the coast of Savannah, GA (vs. 1.7 m observed by NOAA), and 1.6 m along the coast of Charleston, SC (vs. 1.4 m observed by NOAA). We were able to predict these peaks approximately 12 hours in advance of the observed values. Overall, the model had an average error of less than 0.3 m. The storm surge arrived during low tide at these locations, but the situation could have been much worse if the surge occurred during high tide. Our model takes into account the timing of both high/low tide and expected surge.
  • In summary, our large-scale SCHISM storm surge model accurately predicted the potential impacted area in Grand Bahama Island 2 days in advance and the peak storm tides in U.S. Southeast Coasts in particular for forecast tracks #4, #9, and #10 as they correlated closest with the actual path of the storm. The prediction agrees well with later published satellite images of Grand Bahama and the observed peak water levels from NOAA. Our model was able to catch the correct peak in 12–24 hours in advance (Full validation of our model will be conducted in following weeks after NOAA verified water level observations during Dorian and we will publish an update on this once we have that).

    It is well known that the uncertainty in hurricane forecasts as done by NOAA’s NHC is very high at 5–7 days out, which results in associated uncertainty in our storm surge predictions. As a result, it is important to update predictions as we did above once a new hurricane forecast becomes available.

    Methodology: One Concern’s SCHISM Storm Tide Model:

    One Concern’s storm tide model is based on SCHISM (Semi-implicit Cross-scale Hydroscience Integrated System Model) and the model can predict storm tide, a combination of astronomical tide and storm surge, along the entire U.S. East Coast, Gulf of Mexico Coast, and the Bahamas as seen above.

    SCHISM Model

    SCHISM is a derivative work from the original SELFE model and it has been implemented by Dr. Joseph Zhang (College of William & Mary) and other developers around the world (SCHISM wiki). The model has been extensively tested and validated against standard ocean/coastal benchmarks and applied to many coastal regions around the world (See a list of publications here). We chose SCHISM over other storm surge models because of 3 main reasons:

    1. more flexible modeling domain (cross-scale unstructured grids from deep ocean to shallow water)

    2. higher accuracy in numerical solver, and

    3. better representation of ocean bathymetry without smoothing.

    One Concern’s Model Differentiation

    We have made a number of modifications to this base model to create the outputs above. Here are some of our most significant modifications:

    • Developing a high-resolution cross-scale grid which covers the entire U.S. East Coast and Gulf Coast.
    • Building an operational forecast pipeline with at least 150x real-time speedup.
    • Utilizing multiple atmospheric forecasts (GFS, NAM, ECMWF, etc) for probabilistic simulations.
    • Automating model calibration and validation.
    • Seamlessly coupling with inland inundation model. (outputs from this model are not represented above)

    We are also continuing to improve our models in a couple of ways:

    • Due to fully solving the governing equations of momentum and continuity, SCHISM requires much more computation than simplified physics models such as SLOSH. Our engineers are working to make the model more efficient and reduce the computation time.
    • We are also developing new techniques to improve the models’ probabilistic simulations and better quantify the uncertainty associated with the model prediction.

    Conclusion

    In this article we have shown how our model can predict the potential coastal storm tide based on hurricane forecasts for Hurricane Dorian. In the results above, we were able to provide the potential impacted areas in the Grand Bahama Island 2 days in advance and the peak water levels in U.S. Southeast Coast in 12–24 hours in advance prior to the event. For the hurricane forecasts that best match the storm’s path (Predictions #4, #9, and #10 above), our results matched closely with satellite imagery and preliminary tidal gauge observations from NOAA. As our models are inherently based on uncertain hurricane forecast predictions, we are working hard on both model improvements as well as engineering improvements to run models efficiently as soon as new predictions are available and provide uncertainty analysis in the future.