Estimation of radiation scalar using deep learning for improved gross primary productivity estimation based on a light-use efficiency model

Abstract

Accurately quantifying terrestrial gross primary productivity (GPP) can provide insights into the dynamics of global ecosystems. Light Use Efficiency (LUE) models with simple and effective structures have been widely used to estimate GPP. However, GPP estimation using LUE models was biased due to inaccurately representing the nonlinear effect of solar radiation on photosynthesis. To address this issue, we developed a hybrid LUE model, termed LUE-MLP, which integrates a multilayer perceptron (MLP) neural network simulating the relationship between GPP and solar radiation with a big-leaf LUE model. LUE-MLP was optimized and tested for GPP simulation on data from 166 flux sites from the FLUXNET2015 dataset, which covers 10 biome types. Results show that LUE-MLP exhibited a reliable performance (R2 = 0.76 and RMSE = 4.17 μmol m−2 s−1) in estimating (half-)hourly GPP on test data. Compared to the original big-leaf and two-leaf LUE models, LUE-MLP's R² (RMSE) for (half-)hourly GPP simulation across biomes rose (fell) by – 0.01-0.13 (−7.6%−16.8%) and 0.01-0.15 (0.39%−29.1%), respectively. Additionally, LUE-MLP well captured the seasonal variations in tower-based GPP. The proposed model can improve GPP estimation on a global scale, enhancing our understanding of global ecosystem dynamics.

Keywords

• Gross primary productivity
• radiation scalar
• maximum light-use efficiency
• machine learning