Estimation of Leaf Physical and Chemical Parameters Based on Hyperspectral Remote Sensing and Deep Learning Technologies
Ji-bo Yue,Meng-die Leng,Qing-jiu Tian,Wei Guo,Yang Liu,Haikuan Feng,Hong-bo Qiao
DOI: https://doi.org/10.3964/j.issn.1000-0593(2024)10-2873-11
2024-01-01
Spectroscopy and spectral analysis
Abstract:Plant leaf physical and chemical parameters, such as leaf chlorophyll content, carotenoid content, water content, protein content, and Carbone-based constituents content, are crucial for accurately monitoring plant growth status. In recent years, with the rapid development of deep learning technology in vegetation remote sensing, the combined use of deep learning and hyperspectral remote sensing for plant leaf parameters estimation has been widely applied; however, currently, few leaf parameters estimation works based on the combination of deep learning and hyperspectral remote sensing technology have been conducted. This study explores the possibility of estimating leaf chlorophyll, carotenoid, water, protein, and Carbone-based constituent content by combining hyperspectral remote sensing and deep learning techniques. The main work of this paper is to propose a leaf physical and chemical parameter estimation method based on hyperspectral remote sensing and deep learning. Firstly, this study determines the sensitive spectral regions of multiple vegetation leaf physical and chemical parameters based on the PROSPECT-PRO radiative transfer model. Then, we designed a LeafTraitNet deep learning model; the LeafTraitNet model is trained and tested based on the lobex93 dataset, and a high-precision leaf parameter estimation result is obtained. The conclusions of this study are as follows: (1) It is vital to select leaf spectral absorption features based on the PROSPECT-PRO radiative transfer model. The leaf chlorophyll (434 and 676 nm) and carotenoids (445 nm) spectral absorption regions are located in the visible bands. However, the absorption regions with the most significant correlation coefficients (absolute values) are not their maximum spectral absorption bands, which the mutual influence of leaf chlorophyll and carotenoid absorptions may cause. (2) The leaf water spectral absorption regions are mainly located in the bands 950 similar to 2 500 nm, which overlaps with the spectral absorption regions of leaf protein and carbon-based component content, thus weakening the hyperspectral remote sensing estimation accuracy of the latter. The correlation coefficients between leaf protein (and carbon-based component content) and the spectral reflectance in the 950 similar to 2 500 nm range are notably lower than the leaf water. The correlation coefficients analysis results of leaf parameters and hyperspectral for the PROSPECT-PRO radiative transfer model and lobex93 dataset show similar correlation coefficients. (3) The three traditional methods and the LeafTraitNet model can be ranked as LeafTraitNet (total nRMSE=0.84) < RF (total nRMSE=1.59) < MLP (total nRMSE=1. 73) < MLR (total nRMSE=1.74), which means the leaf parameters estimation performance is notably higher than RF, MLP, and MLR. However, further experiments are needed to validate the LeafTraitNet model at the canopy scale.