EGU General Assembly 2023
© Author(s) 2023. This work is distributed under
the Creative Commons Attribution 4.0 License.

Added value of machine learning in the assimilation of ASCAT observations into the ISBA land surface model

Timothée Corchia1, Bertrand Bonan1, Nemesio Rodriguez-Fernandez2, Gabriel Colas1, and Jean-Christophe Calvet1
Timothée Corchia et al.
  • 1CNRM, Université de Toulouse, Météo-France, CNRS, Toulouse, France (
  • 2CESBIO, Université de toulouse, CNES, CNRS, Toulouse, France

In the context of global warming, the frequency and intensity of extreme events such as droughts are increasing, and a better modeling of vegetation response to climate is needed. Monitoring the impact of extreme events on land surfaces involves a number of soil-plant system variables such as soil water content, soil temperature and vegetation leaf area index (LAI). These variables control the carbon, water, and energy land surface fluxes. They can be monitored either by using the unprecedented amount of data from the fleet of Earth observation satellites or by using land surface models. Alternatively, all available sources of information can be combined by assimilating satellite observations into models.


The LDAS-Monde Land Data Assimilation System is a tool developed by the Centre National de Recherches Météorologiques (CNRM). It allows the joint assimilation of Advanced SCATterometer (ASCAT) surface soil moisture and Copernicus Global Land service (CGLS) LAI retrievals into the ISBA (Interaction Sol-Biosphère-Atmosphère) land surface model of Meteo-France, with the objective of better representing leaf biomass and root-zone soil moisture. The ASCAT C-band radar backscatter coefficients (σ0) contain information on both surface soil moisture and vegetation and its assimilation could prove beneficial. For this, an observation operator that links σ0 to the ISBA land surface variables is needed.


In this work, a method for the assimilation of ASCAT σ0 into ISBA using LDAS-Monde is presented. In a first step, observation operators are built using machine learning. Neural networks (NN) are trained using as inputs modeled soil surface moisture, soil temperature, rainwater interception by leaves and CGLS satellite observations of LAI. Then the observation operators are implemented into LDAS-Monde, making it capable of assimilating the satellite product. The method is implemented over southwestern France, where in situ soil moisture observations are available. It is shown that the assimilation of σ0 alone markedly improves the simulation of LAI and soil moisture in agricultural areas. Results vary from one land cover type to another.

How to cite: Corchia, T., Bonan, B., Rodriguez-Fernandez, N., Colas, G., and Calvet, J.-C.: Added value of machine learning in the assimilation of ASCAT observations into the ISBA land surface model, EGU General Assembly 2023, Vienna, Austria, 24–28 Apr 2023, EGU23-6976,, 2023.