From Physical Models to Transfer Learning: Flexible Strategies for Leaf Functional Trait Estimation across Global Optical Properties Datasets
Journal of Remote Sensing · 01.01.2026 · 6
DOI: 10.34133/remotesensing.1050 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Remote Sensing, online erschienen am 07.05.2026. Verfasst von Shuaipeng Fei und 5 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 90 Literaturangaben. Gefördert von: Inner Mongolia Grassland Technology Innovation Center Major Innovation Platform Construction Project, Key Project of Inner Mongolia Science and Technology Promotion Action, Industrial Technology Innovation Program of IMAST.
Abstract (Originalsprache)
Accurate estimation of leaf functional traits is crucial for optimizing crop management, increasing yield, and conserving the ecological environment. This study employed 26 globally collected leaf optical properties datasets. These datasets encompass key leaf functional traits: chlorophyll (CHL), carotenoid content (CAR), equivalent water thickness (EWT), nitrogen content (N), and leaf mass per area (LMA). The optical properties were represented by leaf reflectance spectra (400 to 2,400 nm). For physical modeling, the PROSPECT-D and PROSPECT-PRO models were utilized, incorporating prior information on the leaf structure parameter ( N struct ) and employing optimal spectral domains. In terms of data-driven models, 4 modeling strategies using TabNet, deep residual networks (ResNet), and generalized linear model (GLM) were implemented: (a) training models with small-sample target data, (b) training models with source data from other global datasets, (c) fine-tuning the models trained in strategy (b) on the target data to obtain transfer learning models, and (d) merging source and target data to train models. The hybrid models were constructed using 20,000 PROSPECT-simulated data points to replace source data in strategies (b) to (d). Additionally, decision-level fusion of physical and data-driven models was conducted using 4 fusion algorithms. Results demonstrate that while physical models outperform source-trained models for EWT and LMA, they are consistently surpassed by fine-tuning transfer learning approaches, particularly GLM-based implementations, which achieved peak accuracy. Physical models occasionally exceeded optimal data-driven performance in specific datasets, and hybrid models generally underperformed pure data-driven methods. Based on the consistent performance between target and test datasets, a model selection framework with 97% accuracy was developed, and the estimation accuracy of CHL, CAR, and LMA was improved through Bayesian model averaging (BMA) decision-level fusion. Finally, we developed flexible guidelines for leaf trait estimation, highlighting the essential role of limited manual measurements. These groundbreaking advances provide novel and robust technical support for agricultural management and global change ecology research.
Förderung
- Inner Mongolia Grassland Technology Innovation Center Major Innovation Platform Construction Project
- Key Project of Inner Mongolia Science and Technology Promotion Action
- Industrial Technology Innovation Program of IMAST
Zitieren
APA
Fei, S., Chen, M., Wang, Z., Liu, Y., Xiao, Y., & Ma, Y. (2026). From Physical Models to Transfer Learning: Flexible Strategies for Leaf Functional Trait Estimation across Global Optical Properties Datasets. Journal of Remote Sensing, 6. https://doi.org/10.34133/remotesensing.1050
Vancouver
Fei S, Chen M, Wang Z, Liu Y, Xiao Y, Ma Y. From Physical Models to Transfer Learning: Flexible Strategies for Leaf Functional Trait Estimation across Global Optical Properties Datasets. Journal of Remote Sensing. 2026;6. doi:10.34133/remotesensing.1050
BibTeX
@article{Fei2026,
author = {Fei, Shuaipeng and Chen, Mian and Wang, Zhi and Liu, Yaling and Xiao, Yonggui and Ma, Yuntao},
title = {From Physical Models to Transfer Learning: Flexible Strategies for Leaf Functional Trait Estimation across Global Optical Properties Datasets},
journal = {Journal of Remote Sensing},
year = {2026},
volume = {6},
doi = {10.34133/remotesensing.1050}
}
EUVIMED