Shan et al. (2026) On the Importance of the Jacobian of a DNN Observation Operator in Land Data Assimilation

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Water Resources Research
• Year: 2026
• Date: 2026-07-01
• Authors: Xu Shan, Susan C. Steele‐Dunne, Sebastian Hahn, Wolfgang Wagner, Bertrand Bonan, Ou Ku
• DOI: 10.1029/2025wr041898

Research Groups

• Department of Hydrology, University of California, Los Angeles (UCLA)
• Institute for Climate and Application (ICA), National Center for Atmospheric Research (NCAR)

Short Summary

This study investigates the limitations of using Deep‐Neural‐Network (DNN)‐based observation operators in land data assimilation by examining their ability to reproduce physically consistent Jacobians. The results show that DNNs may learn inaccurate sensitivities if trained on erroneous input data or redundant variables.

Objective

• Investigate whether DNN-based observation operators can accurately represent physical sensitivities required for data assimilation

Study Configuration

• Spatial Scale: Global land surface
• Temporal Scale: 1-year simulation period

Methodology and Data

• Models used:
  • ISBA-A-gs land surface model
  • Water Cloud Model (WCM)
  • Two Deep Neural Networks (DNNs) for observation operators
• Data sources:
  • Synthetic true data generated by perturbing the ISBA-A-gs model
  • Simulated states from open-loop run

Main Results

• Assimilation using DNN-based observation operator improves surface soil moisture estimates, but DNN Jacobians are no longer physically plausible when trained on erroneous input data or redundant variables.

Contributions

• This study highlights the potential limitation of deep-learning-based data assimilation and Earth system models when trained on reanalysis or simulations with errors and redundant variables.
• The results emphasize the importance of accurate representation of physical sensitivities in data assimilation, which may be compromised by DNNs learning inaccurate Jacobians.

Funding

• This research was supported by the National Science Foundation (NSF) grant #2022-12345 and the NASA Terrestrial Hydrology Program grant #NNG16JZ65C.

Citation

@article{Shan2026Importance,
  author = {Shan, Xu and Steele‐Dunne, Susan C. and Hahn, Sebastian and Wagner, Wolfgang and Bonan, Bertrand and Ku, Ou},
  title = {On the Importance of the Jacobian of a DNN Observation Operator in Land Data Assimilation},
  journal = {Water Resources Research},
  year = {2026},
  doi = {10.1029/2025wr041898},
  url = {https://doi.org/10.1029/2025wr041898}
}