COMT Chesapeake Bay Hypoxia Modeling VIMS: Marjy Friedrichs (lead PI) Carl Friedrichs (VIMS-PI) Ike Irby (funded student) Aaron Bever (consultant) Jian Shen (collaborator) WHOI: Malcolm Scully (WHOI-PI) UMCES: Raleigh Hood (UMCES-PI) Hao Wang (funded student) Jeremy Testa (collaborator) Wen Long (collaborator) Meng Xia (collaborator) NOAA-CSDL: Lyon Lanerolle (NOAA-PI) Frank Aikman (collaborator) EPA-CBP: Ping Wang, Lewis Linker (collaborators) NOAA NOS/COOPS Transition Partner: Pat Burke July 30-31, 2015 COMT Annual Meeting

Chesapeake Hypoxia Objective Assess the readiness/maturity of a suite of existing estuarine community models for determining past, present and future hypoxia events within the Chesapeake Bay, in order to accelerate the transition of hypoxia model formulations and products from “academic research” to “operational” centers Chesapeake Hypoxia COMT “operational” centers include: Chesapeake Bay Ecological Prediction System - Quasi-operational short-term forecasts NOAA NOS/CO-OPS - Truly-operational short-term forecasts EPA Chesapeake Bay Program - Regulatory scenario-based forecasts

 3 ROMS-based model variants ChesROMS ROMS-UMCES CBOFS (NOAA operational model)  3 other types of models CH3D (EPA operational/regulatory model) EFDC FVCOM Hydrodynamic Models All hydrodynamic models are similar, but vary in terms of: vertical and horizontal resolution (bathymetry) sigma vs. z-grid structured vs. unstructured grids riverine/atmospheric forcing (in some cases)

4 Full biogeochemistry models ICM (CH3D, FVCOM) RCA (ROMS) ECB (ChesROMS) BGC (ChesROMS) 1 Constant biology models 1term (CBOFS, ChesROMS, EFDC) Dissolved Oxygen (DO) Models  8 model combinations

Outline 1.Model comparisons: How well do these 8 models simulate DO? How can these model simulations be improved? 2. Lessons learned from model comparisons: Uncertainties in computing hypoxic volume by interpolating DO observations Estimating hypoxic volume from a few vertical profilers 3. Interdecadal model simulations: Comparison of two 20-yr simulations: CH3D-ICM & ChesROMS-1term Analysis of 30 year ChesROMS-1term simulation 4. Future Directions: Transitioning to operations

Goals: -Which models best resolve bottom DO in the Chesapeake Bay? -How do modelers move forward in improving model skill of low-oxygen conditions? Chesapeake Hypoxia Model Comparisons

Compare simulations to observations at 13 stations for ~34 total cruises from Jan 2004 – Dec 2005 Chesapeake Hypoxia Model Comparisons

Model Skill Assessment

Target Diagrams Taylor Diagrams bias Unbiased RMSD (STD) Total RMSD Model skill same as skill of mean of observations RMSD = Root mean square difference

Model Skill Assessment Target Diagrams Taylor Diagrams bias Unbiased RMSD

Model Skill Assessment Target Diagrams Taylor Diagrams bias Unbiased RMSD

Model Skill Assessment Target Diagrams Taylor Diagrams bias Unbiased RMSD (STD) Total RMSD

Model Skill Assessment Target Diagrams Taylor Diagrams 1 0 Correlation Coefficient bias Unbiased RMSD (STD) Total RMSD

Model Skill Assessment Target Diagrams Taylor Diagrams 1 0 Correlation Coefficient Standard Deviation bias Unbiased RMSD (STD) Total RMSD

Model Skill Assessment Target Diagrams Taylor Diagrams 1 0 Correlation Coefficient Standard Deviation RMSD bias Unbiased RMSD (STD) Total RMSD

Model Comparison DO at Surface normalized bias normalized unbiased RMSD CH3D-ICM ChesROMS-ECB ChesROMS-BGC ROMS-RCA FVCOM-ICM ChesROMS-1term CBOFS2-1term EFDC-1term Model Mean 1 1 normalized standard deviation RMSD

Model Comparison DO at Surface DO at 5m Depth normalized bias normalized unbiased RMSD CH3D-ICM ChesROMS-ECB ChesROMS-BGC ROMS-RCA FVCOM-ICM ChesROMS-1term CBOFS2-1term EFDC-1term Model Mean 1 1 normalized standard deviation RMSD

Model Comparison DO at Surface DO at 5m Depth DO at 10m Depth normalized bias normalized unbiased RMSD CH3D-ICM ChesROMS-ECB ChesROMS-BGC ROMS-RCA FVCOM-ICM ChesROMS-1term CBOFS2-1term EFDC-1term Model Mean 1 1 normalized standard deviation RMSD

Model Comparison DO at Surface DO at 5m Depth DO at 10m Depth DO at Bottom normalized bias normalized unbiased RMSD CH3D-ICM ChesROMS-ECB ChesROMS-BGC ROMS-RCA FVCOM-ICM ChesROMS-1term CBOFS2-1term EFDC-1term Model Mean 1 1 normalized standard deviation RMSD Overall skill of all models (temporal + spatial variability): High in terms of DO *especially bottom DO

Model Comparison DO at Surface DO at 5m Depth DO at 10m Depth DO at Bottom normalized bias normalized unbiased RMSD CH3D-ICM ChesROMS-ECB ChesROMS-BGC ROMS-RCA FVCOM-ICM ChesROMS-1term CBOFS2-1term EFDC-1term Model Mean 1 1 normalized standard deviation RMSD Overall skill of all models (temporal + spatial variability): High in terms of DO *especially bottom DO

Model Comparison 21 Observation Station CB4.1C 2004 & 2005 Observation Number Model Mean 95% Confidence Interval Observations Models DO Concentration

Model Comparison Temp at Surface Temp at Bottom normalized bias normalized unbiased RMSD CH3D-ICM ChesROMS-ECB ChesROMS-BGC ROMS-RCA FVCOM-ICM ChesROMS-1term CBOFS2-1term EFDC-1term Model Mean 1 1 normalized standard deviation RMSD

Model Comparison Temp at Surface Temp at Bottom Salinity at Surface Salinity at Bottom normalized bias normalized unbiased RMSD CH3D-ICM ChesROMS-ECB ChesROMS-BGC ROMS-RCA FVCOM-ICM ChesROMS-1term CBOFS2-1term EFDC-1term Model Mean 1 1 normalized standard deviation RMSD

Model Comparison normalized bias normalized unbiased RMSD CH3D-ICM ChesROMS-ECB ChesROMS-BGC ROMS-RCA FVCOM-ICM ChesROMS-1term CBOFS2-1term EFDC-1term Model Mean 1 1 normalized standard deviation RMSD Temp at Surface Temp at Bottom Salinity at Surface Salinity at Bottom Chlorophyll at Surface Chlorophyll at Bottom

Model Comparison normalized bias normalized unbiased RMSD CH3D-ICM ChesROMS-ECB ChesROMS-BGC ROMS-RCA FVCOM-ICM ChesROMS-1term CBOFS2-1term EFDC-1term Model Mean 1 1 normalized standard deviation RMSD Temp at Surface Temp at Bottom Salinity at Surface Salinity at Bottom Chlorophyll at Surface Chlorophyll at Bottom Nitrate at Surface Nitrate at Bottom Overall skill of all models (temporal + spatial variability): High in terms of temp and salinity Low in terms of chlorophyll and nitrate

Model Comparison normalized bias normalized unbiased RMSD CH3D-ICM ChesROMS-ECB ChesROMS-BGC ROMS-RCA FVCOM-ICM ChesROMS-1term CBOFS2-1term EFDC-1term Model Mean 1 1 normalized standard deviation RMSD Temp at Surface Temp at Bottom Salinity at Surface Salinity at Bottom Chlorophyll at Surface Chlorophyll at Bottom Nitrate at Surface Nitrate at Bottom DO at Bottom Overall skill of all models (temporal + spatial variability): High in terms of temp and salinity Low in terms of chlorophyll and nitrate

Model Comparison Maximum Stratification & Mixed Layer Depth Depth Surface Bottom Salinity High Low

Model Comparison Maximum Stratification & Mixed Layer Depth Depth Surface Bottom Salinity High Low abs(Surface – Bottom) * 0.10 meter Stratification exists if: Maximum Stratification

Model Comparison Maximum Stratification & Mixed Layer Depth Depth Surface Bottom Salinity High Low abs(Surface – Bottom) * 0.10 meter Stratification exists if: MLD is the depth above the most shallow existence of stratification Maximum Stratification MLD

Model Comparison Maximum Stratification & Mixed Layer Depth Depth Surface Bottom Salinity High Low abs(Surface – Bottom) * 0.10 meter Stratification exists if: MLD is the depth above the most shallow existence of stratification Halocline & Oxycline Maximum Stratification MLD

Model Comparison normalized bias normalized unbiased RMSD CH3D-ICM ChesROMS-ECB ChesROMS-BGC ROMS-RCA FVCOM-ICM ChesROMS-1term CBOFS2-1term EFDC-1term Model Mean 1 1 normalized standard deviation RMSD Salinity MLD Maximum dS/dz

Model Comparison normalized bias normalized unbiased RMSD CH3D-ICM ChesROMS-ECB ChesROMS-BGC ROMS-RCA FVCOM-ICM ChesROMS-1term CBOFS2-1term EFDC-1term Model Mean 1 1 normalized standard deviation RMSD Salinity MLD Maximum dS/dz too shallow too weak

Model Comparison normalized bias normalized unbiased RMSD CH3D-ICM ChesROMS-ECB ChesROMS-BGC ROMS-RCA FVCOM-ICM ChesROMS-1term CBOFS2-1term EFDC-1term Model Mean 1 1 normalized standard deviation RMSD Overall skill of all models (temporal + spatial variability): Underestimate mean and variability of max dx/dz All but CH3D overestimate mean & underestimate variability of MLD Salinity MLD Maximum dS/dz Oxygen MLD Maximum dO/dz too shallow too weak

Observations r 2 = 0.15Stratification defined as > 10% r 2 = 0.47 r 2 = 0.21Stratification defined as > 25% r 2 = 0.81 Maximum dS/dz Maximum dO/dz Salinity MLD (m) Oxygen MLD (m) There is not a strong relationship between dS/dz and dO/dz, but there is a strong relationship between the depth of the halocline and the depth of the oxycline. Ramifications for Habitat Compression.

Observations r 2 = 0.15Stratification defined as > 10% r 2 = 0.47 r 2 = 0.21Stratification defined as > 25% r 2 = 0.81 Maximum dS/dz Maximum dO/dz Salinity MLD (m) Oxygen MLD (m) There is not a strong relationship between dS/dz and dO/dz, but there is a strong relationship between the depth of the halocline and the depth of the oxycline. Ramifications for Habitat Compression.

Model vs Observations r 2 = 0.59Stratification defined as > 10% r 2 = 0.47 Salinity MLD (m) Oxygen MLD (m) Salinity MLD (m) Oxygen MLD (m) The models correctly represent the relationship between the halocline and oxycline ChesROMS-ECB

Model vs Observations r 2 = 0.59Stratification defined as > 10% r 2 = 0.47 Salinity MLD (m) Oxygen MLD (m) Salinity MLD (m) Oxygen MLD (m) The models correctly represent the relationship between the halocline and oxycline. Important the models improve their ability to match the location of the halocline ChesROMS-ECB

All models do similarly well at reproducing bottom DO  Simple oxygen parameterization models can be used for short-term forecasting… BUT, full biogeochemical components will be required for long-term and scenario-based forecasting. The model mean represents the observations better than any single model across the variables examined  Multiple model ensemble can be used to provide confidence bounds of DO forecasts. Observations demonstrate a strong relationship between the salinity MLD and oxygen MLD  Models correctly represent this relationship  Models need to better resolve the location, rather than the degree (as long as stratification exists), of the halocline in order to properly simulate habitat compression from low-oxygen waters. Model Comparison Conclusions

Outline 1.Model comparisons: How well do these 8 models simulate DO? How can these model simulations be improved? 2. Lessons learned from model comparisons: Uncertainties in computing hypoxic volume by interpolating DO observations Estimating hypoxic volume from a few vertical profilers 3. Interdecadal model simulations: Comparison of two 20-yr simulations: CH3D-ICM & ChesROMS-1term Analysis of 30 year ChesROMS-1term simulation 4. Future Directions: Transitioning to operations

Lessons Learned from Model Comparisons Use 3D models to examine uncertainties in EPA observation-based “operational” estimates of Chesapeake Bay hypoxic volume. DO observed by EPA are not a “snapshot” = temporal error DO observed by EPA have coarse spatial resolution = spatial error Use 3D models to improve EPA interpolated estimates of hypoxic volume JGR-Oceans, October 2013 issue:

Temporal Uncertainties in Interpolated Hypoxic Volumes Blue triangles = 13 selected CBP stations  Reduce Temporal errors: 1.Choose subset of 13 CBP stations 2.Routinely sampled within 2.3 days of each other 3.Characterized by high DO variability  This reduces Temporal errors from ~ 5 km 3 to ~ 2.5 km 3

Spatial Uncertainties in Interpolated Hypoxic Volumes  Reduce Spatial errors: Model 13-station cruises (4 models x 2 years x 12 cruises/year) Derive a factor to “correct” 13-station interpolation to equal the Integrated 3D hypoxic volume.  This reduces Spatial errors from ~ 5 km 3 to ~ 2.5 km 3 Before Scaling After Scaling

Spatial Uncertainties in Interpolated Hypoxic Volumes  Reduce Spatial errors: Model 13-station cruises (4 models x 2 years x 12 cruises/year) Derive a factor to “correct” 13-station interpolation to equal the Integrated 3D hypoxic volume.  This reduces Spatial errors from ~ 5 km 3 to ~ 2.5 km 3 Before Scaling After Scaling  Total Temporal + Spatial errors ≈ ~ 2.5 km 3 + ~ 2.5 km 3 ≈ ~ 5 km 3

For a given thickness of the hypoxic layer in the Chesapeake Bay, the horizontal extent of hypoxia is constrained by the steep topography of the Bay’s deep channel. Therefore, the volume of hypoxic water can be reasonably estimated with data from a relatively small number of vertical profiles. If a few automated profilers or well-instrumented moorings can report oxygen data in real time, then the hypoxic volume (HV) of the Bay can be reasonably estimated in real time. Newest work: Can we use a few continuous observations from a few profiles to accurately estimate real-time hypoxic volume? Hypothesis: Lessons Learned from Model Comparisons

1. 3D HV = Sum up volume of each 3D model cell that is hypoxic. This can be done for model output only, not cruise-based observations or profilers. 2. Interpolated HV = Uses Chesapeake Bay Program’s Interpolator software. Bever et al. (2013) showed 13 stations are ideal. Can be done for cruise-based observations (error ≈ +/- ~ 5 km 3 ) but not for profilers. 3. Geometric HV (New method) = Assumes hypoxia is constrained by steep bathymetry and top of hypoxic zone is relatively flat. Can be done with a few profilers. Accuracy can be checked by cruise-based observations and by models. Hypoxic Volume Based on a Few Profilers  Methods: Three ways to estimate hypoxic volume (HV)

 Method s (cont.): Calculating “Geometric HV”:

Hypoxic Volume Based on a Few Profilers  Method s (cont.): Calculating “Geometric HV”:

Hypoxic Volume Based on a Few Profilers  Method s (cont.): Calculating “Geometric HV”:  Results Next: (1) Geometric HV compared to Interpolated HV (2) Geometric HV compared to 3D HV

 Results (1): Compare observed “Geometric HV” with 1 site to observed “Interpolated HV” (13 sites): Hypoxic Volume (HV) based on Monitoring Cruise Data (every 2 to 4 weeks for 28 years)

 Results (1): Compare observed “Geometric HV” with 2 to 3 sites to observed “Interpolated HV” (13 sites): Hypoxic Volume (HV) based on Monitoring Cruise Data (every 2 to 4 weeks for 28 years)

Target diagram indicates that 3 sites for “Geometric HV” are nearly as good as sites 1 = uncertainty in Interpolated HV (± 5 km 3 )  Results (1): Compare observed “Geometric HV” with 1 to 10 sites to observed “Interpolated HV” (13 sites):

 Results (2): Compare modeled “Geometric HV” with 2 sites to “3D HV” from model output (Integrated over 1000s of grid points) daily for 20 years (Model = ChesROMS + 1-term constant net respiration) Best 2 sites 1 = no better than 3D HV mean

 Results (2): Compare “Geometric HV” with 2 to 3 sites to “3D HV” (integrated over 1000s of grid points) daily for 20 years Based on 3D model output, 2 sites for “Geometric HV” are as good as 3.

 Information from multiple models ( ) has been used to assess and reduce uncertainties in present CBP interpolated hypoxic volume estimates 13 stations (sample in 2 days) do as well for HV as or more Temporal and spatial uncertainties: together ~5 km 3  Info from 20+ years based on monitoring and 3D model output suggests that 2 to 3 well-chosen stations can do almost as well as 13 Added error due to 2 to 3 stations is less than uncertainty from 13 2 to 3 automated stations could provide continuous real-time HV  Real-time observations of hypoxia would enhance planned model-based operational DO products and could potentially reduce EPA cruise costs Lessons Learned from Model Comparisons Summary

 Information from multiple models ( ) has been used to assess and reduce uncertainties in present EPA interpolated hypoxic volume estimates for Chesapeake Bay 13 stations (sample in 2 days) do as well for HV as or more Temporal and spatial uncertainties: together ~5 km 3  Info from 20+ years based on monitoring and 3D model output suggests that 2 to 3 well-chosen stations can do almost as well as 13 Added error due to 2 to 3 stations is less than uncertainty from 13 2 to 3 automated stations could provide continuous real-time HV  Real-time observations of hypoxia would enhance planned model-based operational DO products and could potentially reduce EPA cruise costs Lessons Learned from Model Comparisons Summary

Outline 1.Model comparisons: How well do these 8 models simulate DO? How can these model simulations be improved? 2. Lessons learned from model comparisons: Uncertainties in computing hypoxic volume by interpolating DO observations Estimating hypoxic volume from a few vertical profilers 3. Interdecadal model simulations: Comparison of HV from two 20-yr simulations: CH3D-ICM & ChesROMS-1term Analysis of 30 year ChesROMS-1term simulation 4. Future Directions: Transitioning to operations

20-year Hypoxic Volume comparison Interpolated: observations ChesROMS-1term CH3D-ICM based on 13 main stem stations ChesROMS-1term overestimates HV and CH3D-ICM underestimates HV

20-year Hypoxic Volume comparison slope = 0.65 r = 0.73 complex EPA model constant biology model On interannual time scales, constant biology (1-term) model does significantly better than the regulatory model in terms of reproducing our best estimate of hypoxic volume slope = 1.26 r = 0.87

Model-data DO correlation (r) On interannual time scales, ChesROMS constant biology (1-term) model: reproduces hypoxic volume (DO < 2.0 mg/L) better than anoxic volume (DO < 0.2 mg/L) reproduces hypoxic volume better in Jun-Sept than in May  Biological variability is more critical where DO < 0.2 mg/L and earlier in the year Analysis of 30-year ChesROMS-1term simulation May June July Aug Sept < 0.2 mg/L < 2.0 mg/L

Jan-May River Discharge Jan-May Nitrogen Load June-Aug NARR Wind Speed < 2.0 mg/L < 0.2 mg/L Red color denotes p<0.05 Model variability is dominated by wind but also significantly correlated with river discharge. Thus it captures some of the biological variability with no biology. Analysis of 30-year ChesROMS-1term simulation

Jan-May River Discharge Jan-May Nitrogen Load June-Aug NARR Wind Speed < 2.0 mg/L < 0.2 mg/L Jan-May River Discharge Jan-May Nitrogen Load June-Aug NARR Wind Speed June-Aug TPL Wind Speed < 2.0 mg/L < 0.2 mg/L Hypoxia/Anoxia from OBSERVATIONS Red color denotes p<0.05 Model variability is dominated by wind but also significantly correlated with river discharge. Thus it captures some of the biological variability with no biology. Observational variability is dominated by nitrogen loading/river discharge. Observations are not significantly correlated with the NARR wind reanalysis product, but are with observed winds. Analysis of 30-year ChesROMS-1term simulation

Jan-May River Nitrogen Load June-Aug Wind Speed < 2.0 mg/L < 0.2 mg/L Residuals (Model minus Observations) Anoxia residuals are negatively correlated with nitrogen loading  lack of biology explains some of the model-data misfit Hypoxia residuals are more strongly correlated with summer winds  the model is not accurately capturing wind response Is the NARR wind product not good enough? Is the model not responding correctly to the wind forcing? Errors in air T and SST are leading to errors in surface stress Surface flux of O 2 is not accurately represented Red color denotes p<0.05

Outline 1.Model comparisons: How well do these 8 models simulate DO? How can these model simulations be improved? 2. Lessons learned from model comparisons: Uncertainties in computing hypoxic volume by interpolating DO observations Estimating hypoxic volume from a few vertical profilers 3. Interdecadal model simulations: Comparison of HV from two 20-yr simulations: CH3D-ICM & ChesROMS-1term Analysis of 30 year ChesROMS-1term simulation 4. Future Directions: Transitioning to operations

“Operational” Centers/Partners Chesapeake Hypoxia COMT “operational” centers: NOAA NOS/CO-OPS -Truly-operational short-term forecasts Chesapeake Bay Ecological Prediction System (CBEPS) -“Quasi-operational” short-term forecasts EPA Chesapeake Bay Program - Regulatory scenario-based forecasts

Current CBOFS Forecast Baltimore Temperature Baltimore Salinity Time (EDT) 8/5 8/6 8/7 8/8 = observation = nowcast = forecast Temperature Salinity Add Dissolved Oxygen!!!!

Future NOAA NOS/CO-OPS CBOFS Year 3: How do the DO models compare when they are run with operational forcing? - Research version of CBOFS-1term run using NOAA NOS/CO-OPS forcing for 2012, compared with hindcast simulations - Compared with other models also run using operational forcing Years 4 & 5: What is the impact of data assimilation and improved grid/bathymetry on predictive capability of CBOFS- 1term? - Analyze the effect of improvements of CBOFS on DO forecasts

Current CBEPS  CBEPS runs a parallel research operational model to CBOFS; run at UMCES and visualized at VIMS  Provides “quasi-operational” nowcasts and short-term (3-day) forecasts of bottom DO for research, management and public uses in Chesapeake Bay. Nowcast 3day Forecast

Future CBEPS Year 3: Can we quantitatively document a user-base for pseudo-operational model products in CBEPS? - Improve bottom DO product available on pseudo-operational website - Document site usage for various products Years 4 & 5: How will end-users react to improved products? - Add additional model products (HV, model mean) and uncertainties to CBEPS website - Analyze product usage and feedback from users

Future: EPA - CBP Year 3: Will the application of nutrient reduction scenarios to the ROMS-ECB model produce similar changes in DO as those generated by the EPA regulatory model? - Apply reduction strategies used by the EPA to our “research” model - Compare results to those obtained for the EPA model - Of great interest to EPA managers, as our results will help establish uncertainty bounds on the official nutrient reduction requirements Years 4 & 5: Will future climate change inhibit the success of current nutrient reduction requirements? - Utilizing EPA’s climate scenarios, we will examine the potential success of the current nutrient load reduction regulations

Extra Slides