Improving the Spatial and Temporal Estimation of Maize Daytime Net Ecosystem Carbon Exchange Variation Based on Unmanned Aerial Vehicle Multispectral Remote Sensing

Accurate estimation of net ecosystem carbon exchange (NEE) is vital to regional carbon balance. Currently, NEE observations at the canopy scale are mainly based on the chamber method. However, the chamber method is labor intensive, time consuming, and measures only plot-scale NEE. It cannot reflect whole-field NEE with high spatial resolution. In this article, maize daytime NEE variations in four fields under different irrigation treatments in a semiarid area was measured using the chamber method, and the spectral reflectance in the maize canopy at noon was obtained using an unmanned aerial vehicle (UAV) multispectral system. We established a daytime NEE variation estimation model and up-scaled the level of NEE observations in maize canopy using UAV-based remote sensing. Twelve widely used vegetation indices were employed for NEE estimation. To obtain an optimal NEE variation estimation method, we compared the performance of several models, including simple linear regression, multiple stepwise regression, and four machine learning (ML) algorithms. Based on the comparison, the modified triangular vegetation index-2 is the best predictor for analyzing simple linear regression, with a coefficient of determination
R
²
= 0.719. Compared with the simple linear regression, there is no substantial increase in the
R
²
of NEE estimation based on multiple stepwise regression. However, the ML algorithms greatly improved
R
²
values. In particular, the gradient boosting regression model exhibits the best performance (
R
²
= 0.856). This article demonstrates that high-resolution UAV multispectral remote sensing shows great potential in improving the spatial and temporal estimating of maize daytime NEE variations.