A deep learning-based hybrid model of global terrestrial evaporation

Area of Science:

• Earth and Environmental Sciences
• Climate Science
• Hydrology

Background:

• Terrestrial evaporation (E) is crucial for climate but global models simplify transpiration (E_t) constraints due to limited data.
• Existing models often assume linear interactions of factors limiting E_t, overlooking complex environmental influences.

Purpose of the Study:

• To develop a more accurate global model for terrestrial evaporation (E).
• To introduce a novel formulation for transpiration stress (S_t) using deep learning.
• To integrate this S_t formulation into a process-based model for a hybrid global E model.

Main Methods:

• Trained deep learning algorithms using eddy covariance, sap flow, and satellite data.
• Modeled transpiration stress (S_t) as the reduction of E_t from its maximum.
• Embedded the S_t formulation into a process-based E model, creating a bidirectionally coupled hybrid model.

Main Results:

• The hybrid model demonstrated enhanced global estimation of transpiration stress (S_t).
• Comparisons with in situ data and satellite proxies confirmed improved accuracy in global E estimation.
• The new S_t formulation improved the model's ability to capture E dynamics.

Conclusions:

• The developed hybrid model offers a significant advancement in estimating global terrestrial evaporation (E).
• The framework can improve Earth System Models and deepen understanding of this vital climatic variable.
• Deep learning integration provides a powerful tool for modeling complex environmental processes like transpiration stress.