Cornell EWRS Seminar

Transcript

1. Combined climate and hydrologic uncertainties shape projections of future soil moisture across the United States David Lafferty (Cornell BEE) with Ryan Sriver (UIUC) Danielle Grogan & Shan Zuidema (UNH) Klaus Keller & Atieh Alipour (Dartmouth) Iman Haqiqi (Purdue) 0 EWRS Seminar September 19th 2024
2. Soil dynamics play a key role in shaping local hydroclimate and associated risks Soils modulate land-atmosphere interactions across a wide range of spatial and temporal scales 1
3. Soil dynamics play a key role in shaping local hydroclimate and associated risks Soils modulate land-atmosphere interactions across a wide range of spatial and temporal scales 2 Extreme soil moisture conditions can drive risks to many critical systems • Soil moisture affects runoff generation and downstream flood risk • Soil droughts can impact agricultural productivity
4. Soil dynamics play a key role in shaping local hydroclimate and associated risks Soils modulate land-atmosphere interactions across a wide range of spatial and temporal scales 3 Extreme soil moisture conditions can drive risks to many critical systems • Soil moisture affects runoff generation and downstream flood risk • Soil droughts can impact agricultural productivity Ø Accurate, long-term soil moisture simulations are therefore essential for effective water resource management and agricultural planning.
5. Modeling long-term hydrologic changes involves considerable uncertainty 4 Projecting future climate change
6. Modeling long-term hydrologic changes involves considerable uncertainty 5 Projecting future climate change Modeling the hydrologic response
7. Modeling long-term hydrologic changes involves considerable uncertainty 6 Projecting future climate change Modeling the hydrologic response Representation of human actions
8. Modeling long-term hydrologic changes involves considerable uncertainty 7 Projecting future climate change Representation of human actions Modeling the hydrologic response
9. Efforts to quantify these uncertainties are computationally expensive 8 Marshall et al. (2021) Importance of parameter and climate data uncertainty for future changes in boreal hydrology. Water Resources Research Chegwidden et al. (2019) How do modeling decisions affect the spread among hydrologic climate change projections? Exploring a large ensemble of simulations across a diversity of hydroclimates. Earth's Future Karimi et al. (2022) Diagnostic framework for evaluating how parametric uncertainty influences agro-hydrologic model projections of crop yields under climate change. Water Resources Research “Given perennial computational constraints, we might derive more confidence in future projections by emphasizing model diversity…” “we consider selected grid cell sensitivities… as a means of balancing the computational demands of our diagnostic experiments” “Global sensitivity analysis methods… are important for comprehensively characterizing model sensitivity. However, these methods can be computationally expensive.”
10. Our modeling framework aims to address these issues 9 Research goals: 1. Develop a lightweight, fast, and differentiable soil moisture simulation model 2. Calibrate the model against historical soil moisture data, accounting for parameter uncertainty 3. Quantify the relative importance of soil parametric and climate uncertainty in projections of future soil moisture
11. pyWBM: A Python implementation of the WBM soil moisture subroutine 10 WBM: The University of New Hampshire Water Balance Model • Process-based gridded global hydrologic model • Simulates land surface components of the global water cycle • Captures multi-scale interactions including water extraction for use in agriculture and domestic sectors. Grogan et al. (2022) Water balance model (WBM) v.1.0.0: a scalable gridded global hydrologic model with water-tracking functionality, Geosci. Model Dev.
12. pyWBM: A Python implementation of the WBM soil moisture subroutine 11 WBM: The University of New Hampshire Water Balance Model • Process-based gridded global hydrologic model • Simulates land surface components of the global water cycle • Captures multi-scale interactions including water extraction for use in agriculture and domestic sectors. Grogan et al. (2022) Water balance model (WBM) v.1.0.0: a scalable gridded global hydrologic model with water-tracking functionality, Geosci. Model Dev. This work was supported by the U.S. Department of Energy, Office of Science Agreement DE-SC0022141. This work was partially supported by the Thaye Fig. 1: Schematic representation of water fluxes within the simple soil moisture model. Additional details can be found in [1]. [1] Grogan, D.S., et. al., Water balance model (WBM) v.1.0.0: a scalable gridded global hydrologic model with water-tracking functionality, Geosci. Model Dev. (2022). https://doi.org/10.5194/gmd-15-7287-2022 We use a Python-implementation of the soil moisture module within the University of New Hampshire Water Balance Model [1]. Our model is a simple 1-dimensional representation of water fluxes within the active soil layer. Uncertain parameters include: alpha (efficiency of evapotranspiration) beta HBV (efficiency of runoff) wiltingp (wilting point) awCap (water capacity of the active layer) samples for observationa then select parameter c each produc gridpoint-ave The sim reproduce ea product ove domain and ensemble w coverage. S with large dis require str changes. The pre-ca uncertain varying deg wiltingp are applied to t fields from e product and identified. a are allowed t content frac sand, silt, cla poorly identif 4. PARA IDENTIFI Unavailable water Active soil layer Precipitation & snowfall Canopy Evapotranspiration Available water capacity Wilting point Snowpack Runoff pyWBM: Python-based soil moisture subroutine of WBM • Simple bucket model with water storage and flux components • Water inputs include precipitation, snowfall, and snowmelt • Outputs driven by evapotranspiration (Hamon method) • Representation of snowpack, canopy, and runoff • No lateral flow of water; no representation of human actions • Forced by daily precipitation and temperature
13. pyWBM: A Python implementation of the Water Balance Model soil moisture subroutine 12    "& ') ( $'#%    ( !' "$'#%        &'      "& ') !$&"#      !$& !'#      !$!)#      "& ') ' &"#      ' & !'#      ' !)# Fig. 1: Schematic representation of water fluxes within the simple soil moisture model. Additional details can be found in [1]. [1] Grogan, D.S., et. al., Water balance model (WBM) v.1.0.0: a scalable We use a Python-implementation of the soil moisture module within the University of New Hampshire Water Balance Model [1]. Our model is a simple 1-dimensional representation of water fluxes within the active soil layer. Uncertain parameters include: alpha (efficiency of evapotranspiration) beta HBV (efficiency of runoff) wiltingp (wilting point) awCap (water capacity of the active layer) Hypercube sampling to draw 30,000 unique parameter samples for each pseudo- observational product. We then select the best 150 parameter combinations for each product based on the gridpoint-average RMSE. The simple model can reproduce each observational product over most of the domain and can generate an ensemble with reasonable coverage. Specific regions with large disagreements may require structural model changes. The pre-calibration constrains uncertain parameters to varying degrees. awCap and wiltingp are scalar factors applied to the corresponding fields from each observational product and are relatively well- identified. alpha and beta HBV are allowed to co-vary with soil content fractions (percentage 4. PARAMETER IDENTIFICATION Fig. 2: (left) Spatial evaluation of the pre-calibration showing RM coefficient of the best parameter set, and the percentage of observ 150 parameter sets). (right) Daily timeseries for two locations showi set in dark gray, and the top 150 parameter sets in light gray. Associ Fig. 3: Parameter combinations of sufficient skill. The top 150 2. SOIL MOISTURE MODEL Unavailable water Active soil layer Precipitation & snowfall Canopy Evapotranspiration Available water capacity Wilting point Snowpack Runoff pyWBM: Python-based soil moisture submodule of WBM Parameter Relevance Available water capacity (𝑊 !"# ) Determines soil storage capacity Wilting point (𝑊 $# ) Determines minimum soil moisture Drying coefficient (𝛼) Controls evapotranspiration efficiency Runoff shape parameter (𝛽% ) Controls runoff partitioning Potential Evapotranspiration (PET) coefficients Modifies PET throughout the year
14. We encode pyWBM in JAX “JAX is a Python library for accelerator-oriented array computation… designed for high- performance numerical computing and large- scale machine learning” 13 https://jax.readthedocs.io For our use-case, JAX provides: Ø Automatic differentiation Ø Just-in-time compilation Ø Automatic vectorization
15. We encode pyWBM in JAX “JAX is a Python library for accelerator-oriented array computation… designed for high- performance numerical computing and large- scale machine learning” 14 https://jax.readthedocs.io For our use-case, JAX provides: Ø Automatic differentiation Ø Just-in-time compilation Ø Automatic vectorization   
16. Historical soil moisture “observations”: NLDAS-2 and SMAP 15 • In-situ soil moisture observations exhibit limited spatial, temporal, and depth coverage • For calibration, we instead rely on reanalysis-driven model outputs • NLDAS-2 • North American Land Data Assimilation System • Quality-controlled, spatially and temporally consistent, harmonized input data • VIC, NOAH, MOSAIC models • 1/8th degree over North America; hourly, 1979-present day • SMAP • Soil Moisture Active Passive satellite L4 dataset • Assimilated GEOS-5 Catchment model simulation outputs • 9km global; 3-hourly, 2015-present day
17. Historical soil moisture “observations”: NLDAS-2 and SMAP 16 • In-situ soil moisture observations exhibit limited spatial, temporal, and depth coverage • For calibration, we instead rely on reanalysis-driven model outputs • NLDAS-2 • North American Land Data Assimilation System • Quality-controlled, spatially and temporally consistent, harmonized input data • VIC, NOAH, MOSAIC models • 1/8th degree over North America; hourly, 1979-present day • SMAP • Soil Moisture Active Passive satellite L4 dataset • Assimilated GEOS-5 Catchment model simulation outputs • 9km global; 3-hourly, 2015-present day
18. Historical soil moisture “observations”: NLDAS-2 and SMAP 17 • In-situ soil moisture observations exhibit limited spatial, temporal, and depth coverage • For calibration, we instead rely on reanalysis-driven model outputs • NLDAS-2 • North American Land Data Assimilation System • Quality-controlled, spatially and temporally consistent, harmonized input data • VIC, NOAH, MOSAIC models • 1/8th degree over North America; hourly, 1979-present day • SMAP • Soil Moisture Active Passive satellite L4 dataset • Assimilated GEOS-5 Catchment model simulation outputs • 9km global; 3-hourly, 2015-present day
19. We leverage the differentiability of pyWBM to calibrate using gradient descent 18 Ø Parameters are updated in the direction of lower loss
20. We leverage the differentiability of pyWBM to calibrate using gradient descent 19 Ø Parameters are updated in the direction of lower loss • Mini-batch gradient descent • Loss function also includes regularization term • 5 random parameter initializations • No train/validate split o No overfitting was found during validation exercises
21. We leverage the differentiability of pyWBM to calibrate using gradient descent 20 We estimate parameter uncertainty in two ways: 1. Calibrating against 4 target datasets (NLDAS-2 and SMAP) 2. Calibrating using 13 different loss functions Loss function Motivation Anomaly space? Reference RMSE Widely used Yes - MSE Widely used Yes - MAE Used in past soil moisture comparisons Yes Xia et al. (2015) Comparison of NLDAS-2 Simulated and NASMD Observed Daily Soil Moisture. Part I: Comparison and Analysis. J. Hydrometeor. NSE Widely used in hydrology No Gauch et al. (2023) In defense of metrics: Metrics sufficiently encode typical human preferences regarding hydrological model performance. Water Resources Research KGE Widely used in hydrology No Gauch et al. (2023) In defense of metrics: Metrics sufficiently encode typical human preferences regarding hydrological model performance. Water Resources Research Outer 50% RMSE Targets outer quantiles Yes - Outer 20% RMSE Targets outer quantiles Yes - Taylor Skill Score Targets variability and correlation No Xia et al. (2015) Automated Quality Control of In Situ Soil Moisture from the North American Soil Moisture Database Using NLDAS-2 Products. J. Appl. Meteor. Climatol.
22. pyWBM achieves good coverage of the target datasets 21      ( % ( ( % ( ( % ( ( ( %. ) ( % ( ( % ( McLean County, Illinois Ensemble Coverage: 75.45%
23. Climate change projections 22 Ø Downscaled & bias-corrected climate projections from LOCA2 • 6 km resolution over North America • SSP3-7.0 (23 models) and SSP2-4.5 (21 models) Pierce et al. (2023) Future Increases in North American Extreme Precipitation in CMIP6 downscaled with LOCA. J. Hydrometeor Ø Full projection ensemble contains ~2200 unique soil moisture projections • 2 SSP scenarios x ~22 climate models x 52 soil parameters
24. What drives projection uncertainty? Average conditions 23             
25. What drives projection uncertainty? Average conditions 24             
26. What drives projection uncertainty? Average conditions 25 E. Borgonovo, A new uncertainty importance measure, Reliability Engineering & System Safety, (2007)                  
27. What drives projection uncertainty? Extreme dry conditions 26                   E. Borgonovo, A new uncertainty importance measure, Reliability Engineering & System Safety, (2007)
28. The importance of soil parametric uncertainty varies across space 27
29. Trends in soil moisture metrics 28 Projected trends 2023-2099 Delta Sensitivity Analysis
30. Conclusions and future work 29 1. We have developed and calibrated a simple water balance model for simulating soil moisture • We can reproduce more sophisticated model outputs with reasonable accuracy 2. Both climate and hydrologic uncertainties are important for projections of future soil moisture • Soil parameters become dominant when modeling soil moisture extremes
31. Conclusions and future work 30 Simpler simulation models can provide many useful benefits pyWBM (~300 lines of Python) ü Improved uncertainty analysis ü Dynamics more interpretable ü Lower cost of entry ! Reduced realism ! Narrow focus (e.g.) VIC (~18,000 lines of C across 60 files) ü More realistic ü Greater scope and utility ! Large experiments are expensive ! Dynamics often unclear Model complexity Helgeson C, Srikrishnan V, Keller K, Tuana N (2021). Why Simpler Computer Simulation Models Can Be Epistemically Better for Informing Decisions. Philosophy of Science
32. Conclusions and future work 31 Future work: 1. Data release 2. Improved calibration and process representation 3. Crop yield impacts analysis 1. We have developed and calibrated a simple water balance model for simulating soil moisture 2. Both climate and hydrologic uncertainties are important for projections of future soil moisture
33. Thanks! Questions? My email: dcl257@cornell.edu Paper & pyWBM repo: github.com/david0811 EWRS Seminar September 19th 2024 32
34. Supplementary slides 33
35. 34 pyWBM parameters: alpha Grogan, D. S., et al. Water balance model (WBM) v.1.0.0: a scalable gridded global hydrologic model with water-tracking functionality, Geosci. Model Dev. (2022) Larger alpha value means ET occurs more efficiently <latexit sha1_base64="dCT9a0CojaCMs+P5yu8MrjTmQFQ=">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</latexit> g(Ws(t)) = 1 exp[ ↵Ws(t)/Wcap] 1 exp[ ↵]
36. 35 pyWBM parameters: beta Smaller beta value means more runoff is generated <latexit sha1_base64="1qkUPv7+MG77EvlR07W4l/wxaJo=">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</latexit> R(t) = (Pt(t) + Ms(t)) ✓ Ws(t) Wcap ◆ R
37. 36 Hamon method for PET <latexit sha1_base64="yZtlxAxC/FkOgQXNHo1wUJbZIb0=">AAACE3icbVC7SgNBFJ2Nrxhfq5Y2g0GIFmE3kSiIEBTBwiJCXpANYXYySYbMPpi5K4Ql/2Djr9hYKGJrY+ffONmk0MQDA4dzzuXOPW4ouALL+jZSS8srq2vp9czG5tb2jrm7V1dBJCmr0UAEsukSxQT3WQ04CNYMJSOeK1jDHV5P/MYDk4oHfhVGIWt7pO/zHqcEtNQxTyo31Rwc40tcLFr5AnYusHOnx7skoXIQdGJFYKwzHTNr5a0EeJHYM5JFM1Q65pfTDWjkMR+oIEq1bCuEdkwkcCrYOONEioWEDkmftTT1icdUO05uGuMjrXRxL5D6+YAT9fdETDylRp6rkx6BgZr3JuJ/XiuC3nk75n4YAfPpdFEvEhgCPCkId7lkFMRIE0Il13/FdEAkoaBrzOgS7PmTF0m9kLdL+dL9abZ8NasjjQ7QIcohG52hMrpFFVRDFD2iZ/SK3own48V4Nz6m0ZQxm9lHf2B8/gBk3ZrP</latexit> PET(t) = 330.2 ⇤ ⇢sat(t) <latexit sha1_base64="J74w5x42egOyzy7k+ztTdux0jTM=">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</latexit> ⇤ = 1 ⇡ cos 1 [ tan tan ] <latexit sha1_base64="S3iZoW6uUTo9EBaYTXVQuW8ZrEs=">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</latexit> = 23.44 cos  360 365 (N + 10) <latexit sha1_base64="s94f8KxUVxmRXHIb4e51QZbTZKA=">AAACIHicbZDLSgMxFIYzXmu9VV26CRahIpSZqq0gQtGNywq9QWcomTTThmYuJGeEMsyjuPFV3LhQRHf6NKaXhbYeSPj5/3NIzudGgiswzS9jaXlldW09s5Hd3Nre2c3t7TdVGEvKGjQUoWy7RDHBA9YADoK1I8mI7wrWcoe347z1wKTiYVCHUcQcn/QD7nFKQFvdXAXbchB2E0UgxdfY9iShSalolXVwhWvToAAnaVLXNz7FpcpZ0bpIu7m8WTQnhReFNRN5NKtaN/dp90Ia+ywAKohSHcuMwEmIBE4FS7N2rFhE6JD0WUfLgPhMOclkwRQfa6eHvVDqEwCeuL8nEuIrNfJd3ekTGKj5bGz+l3Vi8C6dhAdRDCyg04e8WGAI8ZgW7nHJKIiRFoRKrv+K6YBoRqCZZjUEa37lRdEc4yyW78/z1ZsZjgw6REeogCxUQVV0h2qogSh6RM/oFb0ZT8aL8W58TFuXjNnMAfpTxvcPPjmf4g==</latexit> ⇢sat = 2.167 Psat(t) T(t) + 273.15 <latexit sha1_base64="8qnJNaoFetSETQas1ZgjxQyEt/0=">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</latexit> Psat(t) = 8 < : 0.61078 exp h 17.26939 T (t) T (t)+237.3 i , if T(t) 0 0.61078 exp h 21.87456 T (t) T (t)+265.5 i , if T(t) < 0 Day length: Solar declination: Saturation vapor density: Saturation vapor pressure (Tetens equation):
38. 37 pyWBM parameters: PET coefficients Stefan Siebert & Petra Döll, Quantifying blue and green virtual water contents in global crop production as well as potential production losses without irrigation. Journal of Hydrology (2010) Danielle S. Grogan, Global and regional assessments of unsustainable groundwater use in irrigated agriculture (2016). UNH Doctoral Dissertations.
39. Hindcast-based Sobol sensitivity analysis • Sobol is a global, variance-based sensitivity analysis method that accounts for parameter interactions • At select locations, we sample 81,920 parameter combinations from pre-determined ranges 38 Parameter importance varies by location and output metric!
40. Historical soil moisture “observations”: NLDAS-2 and SMAP 39 Cropland (McLean County, IL) Shrubland (Big Bend, TX) ( . )) . . ( . )) . . ( . )) . . ( . )) . . . Cropland (McLean County, IL) Shrubland (Big Bend, TX)
41. Parameters are allowed to vary spatially as a function of soil and land characteristics 40 Parameter Relevance Spatial Variation Available water capacity (𝑊 !"# ) Determines soil storage capacity Determined empirically from target datasets Wilting point (𝑊 $# ) Determines minimum soil moisture Determined empirically from target datasets Drying coefficient (𝛼) Controls evapotranspiration Varies by soil type Runoff shape parameter (𝛽% ) Controls runoff Varies by soil type and topography PET coefficients Modifies potential evapotranspiration throughout the year Varies by crop and land– use type https://ldas.gsfc.nasa.gov/nldas
42. 41 Auxiliary data • Soil content fractions: o California Soil Resource Lab at UC Davis o Based on STATSGO and SURRGO (USDA) • Land-use fractions: o NLDAS-2 and USDA CDL • Leaf area index: o LAI climatology from Global Land Data Assimilation System o Based on AVHRR satellite imagery
43. 42 Gradient descent • Mini-batch gradient descent o Adam optimizer (learning rate = 0.001) o Batch size = 32 o Minimum 10 epochs, maximum 30 o Hyperparameters are not optimized • Loss function also included regularization term • 5 random parameter initializations • No train/validate split o No overfitting was found during validation exercises o Also see: Shen et al. (2022) Time to Update the Split-Sample Approach in Hydrological Model Calibration
44. 43 . . . . . . . . . . . . . . . . . . . . . .
45. 44  ( '!#" '$!#" !)  !#  !#$  $ )  $ #  $ #$  $ &  #$"$ "  "  " & "  "  "  "  "  " ( "  "  ( #$"$ $$   $$   $$  & $$   $$   $$   $$   $$   $$  ( $$   $$  #$"$" " " &" " " " " "  ( (" " #$"$# "% # "% # "% &# "% # "% # "% # "% # "% # "% (# "% # "% #$"$# )# # )# # )# &# )# # )# # )# # )#  ( # )# # )# (# )# # )# #$"$'$ '$ '$ &'$ '$ '$ '$ '$ '$ ('$ '$ " ! $" (" ! $" " ! $" &"" (&""  ( &"" &"""  (&"""  &"""  % %# (% %# % %# % %#"  (% %#"  % %#"  ( "#$ (( "#$ ( "#$ '  ('  '  ' "## (' "## ' "##  ##%"  ( ( ##%"  ##%"  !#"% ( !#"%  !#"% "## ("## "## "" ("" "" %" (%" %"       
46. McLean County coverage 45      ( % ( ( % ( ( % ( ( ( %. ) ( % ( ( % (      ( % ( ( % ( ( % ( ( ( %. )      ( % ( ( % ( ( % ( ( ( %. ) ( (
47. Big Bend coverage 46 ( % ( ( % ( ( % ( ( ( %. ) ( % ( ( % (    ( % ( ( % ( ( % ( ( ( %. ) ( (    ( % ( ( % ( ( % ( ( ( %. )   
48. pyWBM achieves good coverage of the target datasets 47
49. pyWBM performs best during pairwise comparisons 48           
50. Projection timeseries plots: McLean IL 49
51. Projection timeseries plots: Big Bend TX 50
52. Soil moisture trends: McLean County IL 51
53. 52 SA: soil parameter grouping
54. 53 SA: mid- vs late-century
55. 54 SA: mid- vs late-century
56. 55 SA: mid- vs late-century
57. 56 Delta Moment-Independent Sensitivity Measure <latexit sha1_base64="a1Zni3MRtXeMku7j6dEtTQGDBgc=">AAACMnicbZBNS+RAEIY7fjvrx6hHL42D4B4cEhH1Igwugt4UHRUmQ+h0KjPNdD7oroghk9+0l/0lCx70oIhXf4SdcQ67akHDw/tWUV2vn0qh0bYfrInJqemZ2bn52o+FxaXl+srqlU4yxaHNE5moG59pkCKGNgqUcJMqYJEv4dof/Kr861tQWiTxJeYpdCPWi0UoOEMjefXTC08cugFIZBWEivHCKYudkh57xZ0nSldCiMPQy7fyn9uhV+TUjURAK8sorhK9Pg5pQHOv3rCb9qjoV3DG0CDjOvPqf90g4VkEMXLJtO44dordgikUXEJZczMNKeMD1oOOwZhFoLvF6OSSbholoGGizIuRjtR/JwoWaZ1HvumMGPb1Z68Sv/M6GYYH3ULEaYYQ849FYSYpJrTKjwZCAUeZG2BcCfNXyvvMxIYm5ZoJwfl88le42mk6e829891G62gcxxxZJxtkizhkn7TICTkjbcLJb3JPnsiz9cd6tF6s14/WCWs8s0b+K+vtHcn4qns=</latexit> Si = i = 1 2 Exi fy(y) fy|xi (y) dy E. Borgonovo, A new uncertainty importance measure, Reliability Engineering & System Safety, (2007) • Global SA method • Sensitive to changes across the distribution (compared to variance-based methods) • Accounts for first order interactions