CoSMoS (Coastal Storm Modeling System) Southern California v3.0 projections of coastal cliff retreat due to 21st century sea-level rise

Summary: This dataset contains projections of coastal cliff-retreat rates and positions for future scenarios of sea-level rise (SLR). Projections were made using numerical and statistical models based on field observations such as historical cliff retreat rate, submarine slope, coastal cliff height, and mean annual wave power.
Details: Cliff-retreat rate and position projections are for scenarios of 0, 0.5, 1, 1.5 and 2 meters of sea-level rise (SLR) by the year 2100. Projections were made at CoSMoS cross-shore transects(CST) spaced 100 m alongshore. Generally, projections were not made at transects where the sea cliff was armored or otherwise obstructed (for example, houses on the beach in front of the cliff, road between the beach and cliff), though some local exceptions apply where the obstruction was low enough to be easily overwashed. Spatial projections, such as those in the Google Earth KMZs, were made using a baseline sea-cliff edge from 2010.
Two process-based, numerical cliff-profile evolution models were used to make projections: a soft-rock model by Walkden and Hall (2005, 2011) and a hard-rock model by Trenhaile (2000, 2009, 2011). Both models relate breaking-wave height and period to rock erosion, and distribute erosion vertically over a tidal cycle. Model behavior includes a variable beach slope that varies with the prevailing wave climate, wave run-up (Stockdon and other, 2006), and wave set-up that raises the water level during big-wave events and allows waves to impact the sea cliff with greater efficacy and frequency.
The models were run on idealized cliff profiles extending from about 10 m water depth to 1 kilometer inland from the cliff edge. Profiles were extracted by overlaying the cross-shore transects on a high-resolution digital elevation model (DEM) covering the Southern California study area. Using aerial photography, the presence of a beach was recorded (yes or no) for all transects, and the cliff toe elevation (or beach/cliff junction) was digitized from the DEM profiles. Using historic cliff edge retreat rates by Hapke and Reid (2007), unknown coefficients within the cliff-profile models were calibrated using a Monte Carlo simulation (in other words, coefficients were tuned until the modeled mean retreat rate equaled the observed mean retreat rate for a given transect). This was successful for nearly 1,000 DEM profiles (about 45percent of all cliff transects). For those nearly 1,000 profiles, the profile models were run for two time periods: first, a period of 100–200 years using a historic rate of sea-level rise (2 mm/yr), and then for another 100 years using an accelerated mean rate of sea-level rise (5, 10, 15, or 20 mm/yr). The projected cliff-retreat rates are thus mean annual rates that represent the total retreat that occurred during this second time period.
Each of the resulting nearly 5,000 model runs had one dependent (predicted mean annual cliff retreat for a given sea-level rise scenario) and multiple semi-independent variables that determined the magnitude of the dependent variable (such as historic retreat rate, shore platform slope, cliff height, cliff-toe/beach height, cliff-face slope, mean annual wave power, beach slope). This information was used to train a statistical model called an Artificial Neural Network (ANN). The ANN iteratively maps the independent variables to the dependent variable using linear algebra and a weighting system that gives importance to the variables that most strongly influence future sea-cliff retreat. In the end, the trained ANN is a standalone model that has learned, and can reproduce, the process-based cliff-profile model behavior.
Independent tests between cliff profile model output and ANN output showed very good agreement (R-squared = 0.89 - 0.96; root-mean-square-error less than 0.1 m/yr). The trained ANN was then applied to each cross-shore transect, where observed independent variables such as sea cliff height, historic cliff retreat, mean wave power, shore platform slope, sea-level rise scenario, and cliff toe height were passed through the ANN to yield a prediction of future long-term cliff retreat rate.
Two separate ANNs were trained: one to make predictions of the difference between future and historic cliff retreat rates (in other words, prediction = future cliff-retreat rate*historic cliff-retreat rate) and another to predict the mean trend, or acceleration, of cliff retreat as a function of sea-level rise (in other words, future cliff retreat = m*SLR + historic retreat rate, where the ANN predicts m). Training two separate ANNs allowed for two different predictions for each transect from the same training data.
Of the 2,117 cliff transects, there were 8 for which a prediction could not be made, likely because values of independent variables fell outside of the range used to train the ANNs. A prediction was made for these 8 transects by interpolating from multiple neighboring transects for which results were available.
Uncertainty was tallied using a RMSE approach. The RMSE approach represents cumulative uncertainty from multiple sources and assumes that different sources of error will, at times, cancel each other out. It is therefore not a 'worst-case uncertainty' (in other words, a straight sum of errors) but instead an average uncertainty. Sources of cumulative uncertainty included in the RMSE calculation are the base error of the historic retreat rates that the predictions are relative to (0.2 m/yr; Hapke and Reid, 2007), the difference between ANN and cliff-profile model predictions (about 0.1 m/yr), and the spread between the predictions using the two different ANN models (about 0.1 m/yr). Total RMSE increased with SLR rate and varied between 0.20 and 0.32 m/yr.
Final cliff-retreat rates for a given SLR scenario are an average of the two different ANN predictions. The predictions were nominally smoothed using a Butterworth Filter to increase alongshore continuity and emphasize spatial trends in cliff retreat.
References Cited: Hapke, C.J., and Reid, D., 2007. National Assessment of Shoreline Change, Part 4: Historical Coastal Cliff Retreat along the California Coast: U.S. Geological Survey Open-file Report 2007-1133. http://pubs.usgs.gov/of/2007/1133/
Stockdon, H.F., Holman, R. A., Howd, P. A., Sallenger Jr., A. J., 2006. Empirical parameterization of setup, swash, and runup, Coastal Engineering, Volume 53, Issue 7, Pages 573-588, ISSN 0378-3839, http://dx.doi.org/10.1016/j.coastaleng.2005.12.005.
Trenhaile, A. S., 2000. Modeling the development of wave-cut shore platforms, Marine Geology, Volume 166, Issues 1–4, Pages 163-178, ISSN 0025-3227, http://dx.doi.org/10.1016/S0025-3227(00)00013-X.
Trenhaile, A. S., 2009. Modeling the erosion of cohesive clay coasts, Coastal Engineering, Volume 56, Issue 1, Pages 59-72, ISSN 0378-3839, http://dx.doi.org/10.1016/j.coastaleng.2008.07.001.
Trenhaile, A. S., 2011. Predicting the response of hard and soft rock coasts to changes in sea level and wave height, Climatic Change, Volume 109, Issues 3-4, Pages 599-615, ISSN 0165-0009, http://dx.doi.org/10.1007/s10584-011-0035-7.
Walkden, M. J. A., and Hall, J.W., 2005. A predictive Mesoscale model of the erosion and profile development of soft rock shores, Coastal Engineering, Volume 52, Issue 6, Pages 535-563, ISSN 0378-3839, http://dx.doi.org/10.1016/j.coastaleng.2005.02.005.