Accurately modeling gross primary productivity (GPP) and evapotranspiration (ET) in terrestrial ecosystems is essential for understanding and predicting the global carbon and water cycles. However, current models face considerable uncertainties and limitations when estimating these two core components.

Scientists from the Institute of Applied Ecology of the Chinese Academy of Sciences have made new progress in improving the accuracy of both GPP and ET simulations. The findings, published in Geoscientific Model Development and Journal of Hydrometeorology, are expected to provide methodological support for evaluating carbon and water fluxes across diverse ecosystems at regional scales.

In their study on GPP, the Ecological Climate Research Team led by Dr. WU Jiabing developed the Fast Lightweight Automated Machine Learning (FLAML)-light use efficiency (LUE) model based on the light use efficiency (LUE) approach. LUE models estimate plant productivity by linking absorbed light to photosynthetic output, a key process for carbon fixation. By integrating meteorological data, flux tower observations using the eddy covariance method (a technique that measures exchanges of carbon dioxide, water vapor, and energy between ecosystems and the atmosphere), and satellite-derived indices, the researchers applied FLAML, a lightweight automated machine learning framework, to optimize the model. 

The results showed that FLAML-LUE performed exceptionally well in predicting GPP dynamics, with particularly strong accuracy in mixed and coniferous forests. While its performance was somewhat lower in alpine shrubland and typical grassland, it still outperformed widely used products such as Moderate Resolution Imaging Spectroradiometer (MODIS) and PML. The model demonstrated slight declines in accuracy under extreme conditions of high temperature and vapor pressure deficit, but improved performance was observed for croplands and evergreen broadleaf forests under drought conditions. 

The study highlights the potential of automated machine learning in advancing ecological modeling.

In a parallel study on ET, the team led by Dr. FEI Wenli and Dr. SHEN Lidu evaluated the Noah-MP land surface model in multiple parameterization configurations against multiple ET products across eight land-cover types in the continental United States. ET, which represents the combined process of evaporation from land and transpiration from plants, is a central component of the hydrological cycle. They found that the Noah-MP model, across its multiple parameterization configurations, tended to overestimate ET across most land-cover types, especially in evergreen forests, grasslands, croplands, and barren land. 

In addition, over the course of a year, the model accurately captured the seasonal ET peak in summer but consistently overestimated its intensity. When comparing year-to-year variability, model performance was stronger in arid and semi-arid regions than in humid forests and croplands. The researchers identified the dominant physical processes driving ET. For forests and grasslands, stomatal conductance (the regulation of water loss through plant leaf pores) was the key factor, while in shrublands, savannas, croplands, and bare soil areas, runoff processes played a more significant role. 

These findings highlight the importance of adjusting how models are optimized for different land-cover types to improve ET simulation.

Together, the two studies mark an important step forward in improving terrestrial ecosystem modeling of the carbon and water cycles, offering new methodological insights that can support better climate and ecosystem assessments.