Predicting gross primary productivity of poplar plantations based on solar-induced chlorophyll fluorescence using an improved machine learning model

Gross primary production (GPP) is closely associated with processes such as photosynthesis and transpiration within ecosystems, which is a vital component of the global carbon–water–energy cycle. Accurate prediction of GPP in terrestrial ecosystems is essential for evaluating terrestrial carbon cycle processes. Machine learning (ML) models provide significant technical support in this domain. Presently, there is a deficiency of high-precision and robust GPP prediction variables and models. Challenges such as unclear contributions of predictive variables, extended model training durations, and limited robustness must be addressed. Solar-induced chlorophyll fluorescence (SIF), optimized multilayer perceptron neural networks, and ensemble learning models show the potential to overcome these challenges. This study aimed to develop an optimized multilayer perceptron neural network model and an ensemble learning model, while objectively assessing the capacity of SIF to predict GPP. Identifying robust models capable of enhancing the accuracy of GPP predictions was the ultimate goal. This study utilized continuous observations of SIF and meteorological data collected from 2020 to 2021 ​at a designated research observation station within the Populus plantation ecosystem of the Huanghuaihai agricultural protective forest system in Henan Province, China. By optimizing and evaluating the predictive accuracy and robustness of the models across different temporal scales (half-hourly and daily scales), a multi-layer perceptron (MLP) neural network optimization model based on the back propagation (BP) neural network (BPNN) algorithm (BP/MLP) and MLP and random forest (RF) integration (MLP-RF) ensemble models were constructed, utilizing SIF as the primary predictive variable for GPP. Both the BP/MLP (half-hourly scale model R² ​= ​0.885, daily scale model R² ​= ​0.921) and the MLP-RF (half-hourly scale model R² ​= ​0.845, daily scale model R² ​= ​0.914) models showed superior accuracy compared to the BPNN (half-hourly scale model R² ​= ​0.841, daily scale model R² ​= ​0.918) and the traditional RF (half-hourly scale model R² ​= ​0.798, daily scale model R² ​= ​0.867) models, with the BP/MLP model consistently outperforming the MLP-RF model. The BP/MLP model, which was optimized through particle swarm optimization (PSO), significantly enhanced the robustness of GPP predictions on a half-hourly scale and daily scale. Considering both half-hourly scale and daily scale in the PSO-BP/MLP modeling, the four indicators, light-use efficiency (LUE), photosynthetically active radiation (PAR), absorbed photosynthetically active radiation (APAR), and the variation in SIF with NIR_vP (f_SIF(NIR_vP)), exhibited the potential for enhancing the accuracy of GPP predictions. This study employed a series of model optimization techniques to develop a GPP prediction model with enhanced performance that objectively evaluated the contributions of the predictive variables. This approach provided an innovative and effective method for assessing the carbon cycle in terrestrial ecosystems.