Deep learning-based estimation of plant functional traits from canopy spectral reflectance

Area of Science:

• Plant ecophysiology
• Remote sensing
• Machine learning

Background:

• Accurate quantification of canopy-level plant traits is crucial for agriculture and ecosystem monitoring.
• Traditional spectral analysis methods struggle with complex canopy spectral responses.

Purpose of the Study:

• To develop a novel deep learning framework for enhanced plant trait retrieval from hyperspectral data.
• To address limitations in capturing nonlinear and multivariate spectral characteristics.

Main Methods:

• Proposed the Canopy-level Plant Functional Trait Retrieval Network (CTRN), integrating Kolmogorov-Arnold Networks (KAN), Transformer, and Convolutional Neural Networks (CNN).
• Utilized a comprehensive spectral-trait dataset across diverse species, sensors, and continents for training and evaluation.
• Focused on retrieving ten key canopy functional traits.

Main Results:

• CTRN significantly outperformed existing models, achieving R² > 0.82 for all traits.
• Demonstrated high accuracy (R² ≈ 0.80) for Leaf Mass Area (LMA) and Canopy Chlorophyll content (C) even with limited spectral bands (44 bands at 40 nm resolution).

Conclusions:

• CTRN effectively characterizes complex canopy spectral-trait associations.
• The model offers robust capabilities for accurate plant trait parameter retrieval in ecological and agricultural contexts.