Towards constraining soil and vegetation dynamics in land surface models: Modeling ASCAT backscatter incidence-angle dependence with a Deep Neural Network

Shan, X.; Steele-Dunne, S.C.; Huber, M.; Hahn, Sebastian; Wagner, Wolfgang; Bonan, Bertrand; Albergel, Clement; Calvet, Jean-Christophe; Ku, Ou; Georgievska, Sonja

doi:10.1016/j.rse.2022.113116

Towards constraining soil and vegetation dynamics in land surface models

Title

Towards constraining soil and vegetation dynamics in land surface models: Modeling ASCAT backscatter incidence-angle dependence with a Deep Neural Network

Author

Shan, X. (TU Delft Water Resources)
Steele-Dunne, S.C. (TU Delft Mathematical Geodesy and Positioning)
Huber, M. (TU Delft Water Resources)
Hahn, Sebastian (Technische Universität Wien)
Wagner, Wolfgang (Technische Universität Wien)
Bonan, Bertrand (Université de Toulouse)
Albergel, Clement (Université de Toulouse)
Calvet, Jean-Christophe (Université de Toulouse)
Ku, Ou (Netherlands eScience Center)
Georgievska, Sonja (Netherlands eScience Center)

Date

2022

Abstract

A Deep Neural Network (DNN) is used to estimate the Advanced Scatterometer (ASCAT) C-band microwave normalized backscatter (σ40o), slope (σ′) and curvature (σ″) over France. The Interactions between Soil, Biosphere and Atmosphere (ISBA) land surface model (LSM) is used to produce land surface variables (LSVs) that are input to the DNN. The DNN is trained to simulate σ40o, σ′ and σ″ from 2007 to 2016. The predictive skill of the DNN is evaluated during an independent validation period from 2017 to 2019. Normalized sensitivity coefficients (NSCs) are computed to study the sensitivity of ASCAT observables to changes in LSVs as a function of time and space. Model performance yields a near-zeros bias in σ40o and σ′. The domain-averaged values of ρ are 0.84 and 0.85 for σ40o and σ′, compared to 0.58 for σ″. The domain-averaged unbiased RMSE is 8.6% of the dynamic range for σ40o and 13% for σ′, with land cover having some impact on model performance. NSC results show that the DNN-based model could reproduce the physical response of ASCAT observables to changes in LSVs. Results indicated that σ40o is sensitive to surface soil moisture and LAI and that these sensitivities vary with time, and are highly dependent on land cover type. The σ′ was shown to be sensitive to LAI, but also to root zone soil moisture due to the dependence of vegetation water content on soil moisture. The DNN could potentially serve as an observation operator in data assimilation to constrain soil and vegetation water dynamics in LSMs.

Subject

ASCAT
Scatterometry
Radar
Vegetation
Land surface model
Machine learning
Deep Neural Network
Plant water dynamics
Soil moisture

To reference this document use:

http://resolver.tudelft.nl/uuid:3a090606-10b1-432f-ac40-dbc9396360ac

DOI

https://doi.org/10.1016/j.rse.2022.113116

ISSN

0034-4257

Source

Remote Sensing of Environment: an interdisciplinary journal, 279

Part of collection

Institutional Repository