Estimating global terrestrial water storage components by a physically constrained
recurrent neural network$^*$

$^*$Kraft et al. (2022), Towards hybrid modeling of the global hydrological cycle hess-26-1579-2022
Basil Kraft¹ ² (bkraft@bgc-jena.mpg.de), Martin Jung¹, Marco Körner², Sujan Koirala¹, Markus Reichstein¹
1) MPI for Biogeochemistry, Jena, Germany
2) TUM Dept. Aerospace and Geodesy, Munich, Germany
Session HS2.5.2: Recent advancement in estimating global, continental and regional scale water balance components
Global Diagnostic Modelling Group
Max Planck Institute for Biogeochemistry 		Computer Vision Research Group
Technical University of Munich 		European Research Council (ERC) Synergy Grant
Understanding and modeling of the Earth system with ML

Estimating water storages:
a major challenge in hydrology

Water storage components and major land surface processes
Image source g3p website; climers.geo.tuwien.ac.at/climers/research/soil-moisture/g3p/

Still large uncertainties in global
hydrological models

Seasonal soil moisture anomalies.
Image source: Schellekens (2017), doi.org/10.5194/essd-9-389-2017

Still large uncertainties in global
hydrological models

  • Complementary approaches needed
  • Make use of available EO data, e.g.,
    • total water storage variations
    • snow water equivalent
    • evapotranspiration
    • gridded runoff

Why more “data driven”?

  • Knowledge may be wrong or incomplete
    $\rightarrow$ biases
  • We have large amounts of Earth observation data
    $\rightarrow$ we can learn from this data
  • Data driven not better but different
    $\rightarrow$new and complementary insights

Hybrid modeling; learning from data

Hybrid modeling combines machine learning and physically based modeling. B. Kraft

Hybrid modeling: global and spatiotemporal parameters

Hybrid modeling combines machine learning and physically based modeling. B. Kraft
A dynamic hybrid model that combines a recurrent neural network and a
dynamic physically based model. B. Kraft
A dynamic hybrid model that combines a recurrent neural network and a
dynamic physically based model. B. Kraft
A dynamic hybrid model that combines a recurrent neural network and a
dynamic physically based model. B. Kraft

Learned water partitioning fractions
$\ $
$\ $
$\ $
$\ $
$\ $

Cummulative water deficit (CWD) vs water input partitioning fractions. Larger CWD values mean drier soil.
Total water storage decomposition into snow (SWE), soil moisure (SM), and groundwater (GW). Top row: hybrid model (H2M), bottom: physically-based model (PCR-GLOBWB). The seasonality (MSC) and anomalies (IAV) are shown.

Next steps (work in progress)

  • Uncertainty-aware model
  • Identify equifinalities
  • Use additional constraints
    • increase model complexity
    • use additional data constraints

Paper doi.org/10.5194/hess-26-1579-2022

Code github.com/bask0/h2m

Simulations dx.doi.org/10.17617/3.65

bkraft@bgc-jena.mpg.de
@BasilKraft

Model performance

Model performance of the hybrid model (H2M) and four process-based models. Diamond markers (♦️) represent globally averaged signal, bars the cell-level performance.

Hybrid modeling: equifinality

Simulated water storages
$\ $
$\ $
$\ $
$\ $
$\ $

Learned water partitioning fractions
$\ $
$\ $
$\ $
$\ $
$\ $

Self-paced multi-task learning

$\sigma_i$: Task uncertainty, trainable parameter

Kendall, A., Gal, Y. and Cipolla, R., 2018. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 7482-7491).