Coupling a light use efficiency model with a machine learning-based water constraint for predicting grassland gross primary production

Light use efficiency (LUE) model was established to estimate gross primary production (GPP) for understanding the carbon-climate feedbacks of the terrestrial ecosystems. However, water constraints in LUE models can cause large uncertainties in GPP estimates, especially in semiarid grasslands where water is a key forcing factor for multiple ecosystem processes. Here, we proposed a novel LUE-Gradient Boosting Regression Trees (GBRT) model framework where water scalar is derived from five different water constraints for improving the estimates of grassland GPP over the conterminous United States (CONUS). The performance of LUE-GBRT and ten other GPP models [i.e., LUE-RF, LUE-ERT, GBRT, RF, ERT, LUE-fEF, LUE-fVPD, LUE-fLSWI, LUE-fSM, and LUE-fLST] was evaluated against data from eddy covariance (EC) observations at 25 measurement sites over the CONUS domain from 2000 to 2021. We found that LUE-GBRT improved grassland GPP estimates at all EC sites and yielded the highest Kling-Gupta efficiency (0.85) and the lowest root-mean-square error (1.4 g C m- 2 d-1) when compared with the five individual GPP models. LUE-GBRT also showed a superior performance compared to LUE-RF and LUE-ERT. Compared with GBRT, the improvements were particularly from the responses to extreme surface conditions that were better characterized and estimated. An innovation of this method is that LUE-GBRT takes machine learning complementary to the physical-based LUE framework for an optimal junction between GPP physical process and model accuracy.

Automated summary

A novel LUE-Gradient Boosting Regression Trees (GBRT) model framework was proposed to improve the estimation of grassland GPP by deriving water scalar from five different water constraints. The performance of the LUE-GBRT model was evaluated and compared with other GPP models, and it was found that LUE-GBRT showed superior performance with the highest Kling-Gupta efficiency and lowest root-mean-square error. The innovation of this method lies in its combination of machine learning and physical-based framework for optimal accuracy.