Slide 1

Slide 1 text

2018/10/10 9th Japanese Data Assimilation Workshop Hydrometeorology as an inversion problem: Can river discharge observations improve the atmosphere by ensemble data assimilation? Yohei Sawada1,2, Tosiyuki Nakaegawa1, Takemasa Miyoshi2, Tomoki Ushiyama3 1: Meteorological Research Institute, Japan Meteorological Agency 2: RIKEN Advanced Institute for Computational Science 3: International Centre for Water Hazard and Risk Management

Slide 2

Slide 2 text

2018/10/10 9th Japanese Data Assimilation Workshop 1. Introduction

Slide 3

Slide 3 text

2018/10/10 9th Japanese Data Assimilation Workshop 1.1 Disaster prediction in hydrometeorology Grand Challenge: The seamless prediction of floods and droughts observation Atmospheric model Forcings to land surface (e.g., rainfall and radiation) Land Surface Model Hydrological Model observation Disaster monitoring and prediction meteorology hydrology

Slide 4

Slide 4 text

2018/10/10 9th Japanese Data Assimilation Workshop 1.1.1 Example: Global flood prediction obs TIGGE multimodel [Bougeault et al, 2010] forcings HTESSEL[Balsamo et al., 2009 HESS] CaMa-Flood[Yamazaki et al., 2011 WRR] obs Global flood prediction [Zsoter et al., 2016 JHM]  Global flood prediction was achieved by combining atmospheric multmodel ensemble, land surface model, and hydrodynamic model.

Slide 5

Slide 5 text

2018/10/10 9th Japanese Data Assimilation Workshop 1.1.2 Drought prediction obs GFDL seasonal forecast forcings EcoHydro-SiB [Sawada and Koike 2014; Sawada et al. 2015] obs [Sawada and Koike, 2016 JGR] Biomass prediction blue:Forecast gray:climatorogy green: Horn of Africa drought (2010-2011) rainfall intercept evaporation soil moisture transfer ground water photosynthesis growth

Slide 6

Slide 6 text

2018/10/10 9th Japanese Data Assimilation Workshop 1.3 Limitation of traditional framework observation Atmospheric model Forcings to land surface (e.g., rainfall and radiation) Land Surface Model Hydrological Model observation Disaster monitoring and prediction meteorology hydrology

Slide 7

Slide 7 text

2018/10/10 9th Japanese Data Assimilation Workshop 1.4 Hydrology as an inversion problem observation Atmospheric model Forcings to land surface (e.g., rainfall and radiation) Land Surface Model Hydrological Model observation Disaster monitoring and prediction meteorology hydrology  By assimilating land surface hydrological observations, meteorological forcings can be analyzed.

Slide 8

Slide 8 text

2018/10/10 9th Japanese Data Assimilation Workshop 1.4.1 Example: From river discharge to rainfall obs Atmospheric model rainfall Lumped rainfall-runoff model [Boyle, 2000] River discharge [Vrugt et al., 2008 WRR]  By assimilating river discharge observations into a simple rainfall-runoff model, observation rainfall is corrected. [See also Herrnegger et al. 2015 HESS]

Slide 9

Slide 9 text

2018/10/10 9th Japanese Data Assimilation Workshop 1.4.2 Example: From soil moisture to rainfall obs Atmospheric model rainfall Land surface model Satellite Soil moisture  By assimilating AMSR-E soil moisture observation, TRMM precipitation is corrected. [Crow et al., 2011 WRR]

Slide 10

Slide 10 text

2018/10/10 9th Japanese Data Assimilation Workshop 1.5 Is hydrometeorology as an inversion problem possible? observation Atmospheric model Forcings to land surface (e.g., rainfall and radiation) Land Surface Model Hydrological Model observation Disaster monitoring and prediction meteorology hydrology Inverse hydrology estimates forcings to land surface from land hydrological obs. Inverse hydrometeorology may be able to estimate other atmospheric state variables (e.g., wind and water vapor) from land hydrological obs.

Slide 11

Slide 11 text

2018/10/10 9th Japanese Data Assimilation Workshop 1.6 Weakly vs strongly coupled data assimilation Weakly coupled DA Model A domain Model B domain obs obs  Weakly coupled DA performs the analysis update separately for each model domain. Strongly coupled DA Model A domain Model B domain obs Atmosphere River  Strongly coupled DA enables observations in one model domain to directly impact the variables in the other model domain using cross-domain correlations

Slide 12

Slide 12 text

2018/10/10 9th Japanese Data Assimilation Workshop 1.7 Strongly coupled DA is actively investigated in Ocean-Atmosphere DA [Sluka et al., 2016 GRL]  Strongly coupled ocean-atmosphere DA has been successfully done by LETKF.

Slide 13

Slide 13 text

2018/10/10 9th Japanese Data Assimilation Workshop 1.8 Goals • Developing a strongly coupled river-atmosphere data assimilation system to assimilate river discharge observations into the atmosphere. • Performing a proof-of-concept observing system simulation experiment (OSSE) to evaluate the possibility to improve the skill of simulating heavy rainfalls

Slide 14

Slide 14 text

2018/10/10 9th Japanese Data Assimilation Workshop 2. Potential of strongly coupled river-atmosphere data assimilation

Slide 15

Slide 15 text

2018/10/10 9th Japanese Data Assimilation Workshop 2.1 Models & experiment design Models Japan Meteorological Agency’s Non-Hydrostatic atmospheric Model (JMA-NHM) [Saito et al., 2006, 2007] • 3-ice bulk cloud microphysics • Mellor and Yamada level 3 • No convective parameterization is used Lumped 3-layer tank model [Ishihara and Kobatake, 1979] • Operationally used by JMA to assess the landslide risks • Mimicking surface and subsurface flow process by simple tanks • A river basin is lumped as an one grid. rainfall [Nakaegawa et al., 2014 JSCE]

Slide 16

Slide 16 text

2018/10/10 9th Japanese Data Assimilation Workshop 2.1 Models & experiment design Models Japan Meteorological Agency’s Non-Hydrostatic atmospheric Model (JMA-NHM) [Saito et al., 2006, 2007] • 3-ice bulk cloud microphysics • Mellor and Yamada level 3 • No convective parameterization is used Lumped 3-layer tank model [Ishihara and Kobatake, 1979] • Operationally used by JMA to assess the landslide risks • Mimicking surface and subsurface flow process by simple tanks • A river basin is lumped as an one grid. rainfall Experiment Design • 1km horizontal grid spacing • 50 vertical levels with a 22km model top • 36-h ensemble forecast of a heavy rainfall event in September 2015 in Kanto region (78 members) Upper Kinu river [MLIT HP]

Slide 17

Slide 17 text

2018/10/10 9th Japanese Data Assimilation Workshop 2.2 Ensemble forecast of heavy rainfall and river discharge Simulated river discharge  The uncertainty in a forecast of heavy rainfall system’s locations strongly degrades reliability of flood forecasting. 1-h accumulated rainfall [Sawada et al. 2018 JGR]

Slide 18

Slide 18 text

2018/10/10 9th Japanese Data Assimilation Workshop 2.3 Cross-domain correlation between wind field and river discharge Cross-domain (linear) correlation is a key for strongly coupled data assimilation @Sep 9 15UTC Wind field at 925m Southerly moisture inflow Correlation with river discharge Zonal wind Meridional wind  There are meaningful correlations between the wind speed and river discharge in our coupled model. [Sawada et al. 2018 JGR]

Slide 19

Slide 19 text

2018/10/10 9th Japanese Data Assimilation Workshop 2.4 Is background error Gaussian? @Sep 9 21UTC Histograms among 78 ensemble members The Gaussian probabilistic distribution function of a background error is assumed in many conventional data assimilation methods (e.g., LETKF). Rainfall in a single grid Rainfall is not Gaussian Rainfall DA is not straightforward [Koizumi et al. 2005 SOLA; Kotsuki et al. 2017 JGR] Basin-averaged rainfall Spatially-averaged rainfall tends to be Gaussian River discharge River discharge (~temporally averaged basin-averaged rainfall) tends to be Gaussian  Strongly coupled river-atmosphere data assimilation is promising and easy to implement. [Sawada et al. 2018 JGR]

Slide 20

Slide 20 text

2018/10/10 9th Japanese Data Assimilation Workshop 3. A proof-of-concept OSSE

Slide 21

Slide 21 text

2018/10/10 9th Japanese Data Assimilation Workshop 3.1 Strongly coupled river-atmosphere data assimilation system [Sawada et al. 2018 JGR]  Using the cross-domain error covariance, our data assimilation system enables river discharge observations in the rainfall-runoff domain to directly impact the variables in the atmospheric domain during the analysis update.

Slide 22

Slide 22 text

2018/10/10 9th Japanese Data Assimilation Workshop 3.2 Design of OSSE 78 ensemble forecasts Nature run (an ensemble with the 2nd largest river discharge) “Observations” is generated with observation error = 100 [m3/s] Data assimilation with 77 members • 4D-LETKF • Hourly observation • 6h window • No localization Initial & boundary conditions are shared

Slide 23

Slide 23 text

2018/10/10 9th Japanese Data Assimilation Workshop 3.3.1 Skill to simulate meteorological variables (1) Time-mean background RMSE difference between DA and NoDA improve degrade Zonal wind at 925m improve degrade Meridional wind at 925m Humidity at 925m improve degrade Rainfall improve degrade [m/s] [m/s] [g/kg] [mm/h]  We can improve the simulation of the atmospheric state variables as well as rainfall around the target river basin.

Slide 24

Slide 24 text

2018/10/10 9th Japanese Data Assimilation Workshop 3.3.1 Skill to simulate meteorological variables (2) Horizontal wind at 925m Meridional wind at 925m humidity at 925m Rainfall Time series of RMSE difference between DA and NoDA

Slide 25

Slide 25 text

2018/10/10 9th Japanese Data Assimilation Workshop 3.4 4-h forecast at Sep 9 12UTC 4-h accumulated rainfall Truth Change by DA Improvement improve degrade Wind & total condensate water mixing ratio improve degrade Truth Change by DA Improvement  We can improve the location of the entire convection system so that our method is useful for severe rainfall forecasts. [mm] [mm] [mm] [g/kg] [g/kg] [g/kg]

Slide 26

Slide 26 text

2018/10/10 9th Japanese Data Assimilation Workshop 4. Conclusion and Discussions

Slide 27

Slide 27 text

2018/10/10 9th Japanese Data Assimilation Workshop 4.1.1 Discussions (1) • The setting of our OSSE is optimistic. The applicability of river-atmosphere DA should be thoroughly evaluated in the real-world application • In this OSSE, the impact of river obs  atm state was quantified. • How about atm obs  river state? Fully-coupled DA is more promising! • Complete form of strongly coupled DA with the sophisticated runoff-inundation model is new being developed!! Atmospheric and river observations are exchanged! Runoff-inundation modeling with 100m resolution [Sayama et al. 2012; Ushiyama et al. 2016]

Slide 28

Slide 28 text

2018/10/10 9th Japanese Data Assimilation Workshop 4.1.2 Discussions (2) • Observation error in river discharge? • The overall error was the range from 6.2% to 42.8% [Di Baldassarre and Montanari 2009 HESS], but strongly depends on river characteristics • Appropriate localization is needed • The time-lag between rainfall and runoff is important • No hope to improve rainfall at the upstream of Amazon by assimilating river discharge at the Amazon river mouth. • Human impacts (e.g., dam operations and irrigations)

Slide 29

Slide 29 text

2018/10/10 9th Japanese Data Assimilation Workshop 4.2 Conclusions observation Atmospheric model Forcings to land surface (e.g., rainfall and radiation) Land Surface Model Hydrological Model observation Disaster monitoring and prediction meteorology hydrology River discharge Soil moisture Vegetation Groundwater Surface temperature Let’s do hydrometeorology backward!! Ultimate goal: Strongly coupled Earth System DA is possible?

Slide 30

Slide 30 text

2018/10/10 9th Japanese Data Assimilation Workshop JMA-NHM RRI model rain Obsop (atmosphere) LETKF Obsop (river) Atm obs River discharge obs