A deep learning-based hybrid model of global terrestrial evaporation

Abstract

Terrestrial evaporation (E) is a key climatic variable that is controlled by a plethora of environmental factors. The constraints that modulate the evaporation from plant leaves (or transpiration, E_t) are particularly complex, yet are often assumed to interact linearly in global models due to our limited knowledge based on local studies. Here, we train deep learning algorithms using eddy covariance and sap flow data together with satellite observations, aiming to model transpiration stress (S_t), i.e., the reduction of E_t from its theoretical maximum. Then, we embed the new S_t formulation within a process-based model of E to yield a global hybrid E model. In this hybrid model, the S_t formulation is bidirectionally coupled to the host model at daily timescales. Comparisons against in situ data and satellite-based proxies demonstrate an enhanced ability to estimate S_t and E globally. The proposed framework may be extended to improve the estimation of E in Earth System Models and enhance our understanding of this crucial climatic variable.

Fig. 1. Hybrid model architecture.

Schematic of the hybrid terrestrial evaporation model, including the representation sub-grid heterogeneity and the difference in the footprints of the deep learning model and the hybrid model. E_i is interception, E_p is potential evaporation, S is the evaporative stress factor, S_t is transpiration stress, E is actual evaporation, P is precipitation, R_n is net radiation, T_a is air temperature, VOD is vegetation optical depth, VPD is vapor pressure deficit, SW_i is incoming shortwave radiation, and CO₂ is carbon dioxide. The red arrows indicate modeling steps which are exclusive to the processed-based model, the green arrows are steps which have been added in the hybrid, and the black arrows are steps common to both the models.

Fig. 2. Summary of in situ validation of the hybrid and process-based models.

a, b Violin plots showing the distribution of the Kling-Gupta Efficiency (KGE) metric for the transpiration stress factor (S_t) and evaporation (E), respectively, calculated for all flux tower and sap flow measurement sites. The KGE distribution for the hybrid and process-based models are classified according to short and tall vegetation types. The dashed lines represent the median (large dashes) and the interquartile range (small dashes). The red line represents a KGE value of -0.41, above which a model prediction or simulation is considered better than the mean seasonal cycle. For the sap flow sites, transpiration estimates (E_t) instead of E are used.

Fig. 3. In situ comparison of the hybrid and process-based models.

Maps showing the difference in the Kling-Gupta Efficiency (KGE) metric between the hybrid model and process-based model for the transpiration stress factor (S_t) and evaporation (E) calculated using observations at flux tower and sap flow measurement sites in different geographical zones: North America (NA), Asia (AS), Europe (EU), Rest of the World (RW). Blue (red) tones indicate an improvement (degradation) in the hybrid model compared to the process-based counterpart. For the sap flow sites, transpiration estimates (E_t) instead of E are used.

Fig. 4. Global scale evaluation of modeled transpiration stress.

Comparison of the seasonal mean transpiration stress factor (S_t) from the processed-based and hybrid models and the ratio of solar-induced chlorophyll fluorescence and photosynthetically-active radiation (SIF/PAR) for June-July-August (JJA) a, c, e and December-January-February (DJF) b, d, f seasons. Note: The unit of measurement of SIF is mWm²/sr/nm whereas PAR is in W/m².

Fig. 5. Global scale evaluation of modeled evaporation.

Comparison of the seasonal aggregates of evaporation (E) from the processed-based and hybrid models compared with a purely machine learning-based model trained directly on evaporation from FLUXNET sites as the target variable (FLUXCOM) for JJA a, c, e and DJF b, d, f seasons. Note: The units of E is mm/month.

Fig. 6. Correlation of modeled transpiration stress and evaporation with global datasets.

a, c Comparison of correlation maps for transpiration stress factor (S_t) between processed-based and hybrid models with observational SIF/PAR. b and d, Comparison of correlation maps for evaporation (E) between processed-based and hybrid models with machine learning-based estimates (FLUXCOM). e Difference between a, c. f Difference between b and d.