Intelligent retrieval of leaf traits using hyperspectral reflectance and deep learning

Abstract

Reliable and intelligent retrieval of leaf traits from hyperspectral reflectance is crucial for assessing ecosystem functions, yet conventional approaches struggle with spectral complexity and nonlinearities. To address these challenges, we developed the Leaf Trait Retrieval Network (LTRN), a novel deep learning framework that integrates Kolmogorov–Arnold Network (KAN), Transformer, and Temporal Convolutional Networks (TCN) for end-to-end trait estimation. Model validation was carried out using a large spectral–trait database covering hundreds of plant species and four functional traits. Experimental results demonstrated that LTRN model outperforms state-of-the-art deep learning models, achieving R² values greater than 0.78 for estimating chlorophyll content (Chlₐ₊b), equivalent water thickness (EWT), carotenoid content (Ccₐᵣ), and leaf mass per area (LMA). Further analyses indicated that the LTRN model delivers stable estimation performance across spectral resolutions of 10–25 nm. Moreover, the model demonstrates strong stability across varying proportions of training samples. These findings underscore the robustness and stability of LTRN for large-scale vegetation trait retrieval, offering a valuable framework for advancing the intelligent estimation of other ecological parameters.

Plant phenotyping relevance

植物のハイパースペクトル反射から葉形質を推定する深層学習手法を開発し、複数形質・種を含むデータベースで性能と頑健性を検証しており、表現型取得・抽出法が中心である。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.