Estimating leaf functional traits from reflectance spectra is fundamental to remote sensing applications, including precision agriculture, forest monitoring, and climate change studies. While physical radiative transfer models (e.g., PROSPECT) provide a mechanism-driven framework, their accuracy is limited by spectral variability and structural uncertainties. Conversely, data-driven machine learning models excel at pattern recognition but often require abundant labeled data—a resource that is not always available across diverse ecosystems and sensor platforms. Hybrid approaches have emerged to bridge these two paradigms, yet their relative performance and generalizability remain underexplored. Based on these challenges, an in-depth comparative investigation is urgently needed to establish flexible, data-efficient strategies for accurate and transferable leaf trait estimation across global optical properties datasets.
Published (DOI: 10.34133/remotesensing.1050) on May 7, 2026, in the Journal of Remote Sensing (Volume 6, Article ID: 1050), a collaborative research team from China Agricultural University, the Chinese Academy of Agricultural Sciences, and the Inner Mongolia Pratacultural Technology Innovation Center addressed a critical challenge in remote sensing: reliably estimating key leaf functional traits—Leaf functional traits—chlorophyll (CHL), carotenoid (CAR), equivalent water thickness (EWT), nitrogen (N), and leaf mass per area (LMA)—from leaf reflectance spectra (400–2400 nm). Accurate estimation of these traits is crucial for optimizing crop management, increasing yield, and conserving the ecological environment, yet traditional approaches struggle with data scarcity and spectral variability across global ecosystems. This study delivers a comprehensive solution by comparing multiple modeling strategies and developing flexible selection guidelines.
The study systematically compared physical models (PROSPECT-D and PROSPECT-PRO), data-driven models (TabNet, ResNet, and generalized linear model, GLM), hybrid models (incorporating 20,000 PROSPECT-simulated points), and decision-level fusion across 26 global datasets. The most significant finding is that fine-tuning transfer learning approaches—particularly GLM-based implementations—achieved peak accuracy for CHL, CAR, N, and EWT estimation, consistently surpassing physical models. Notably, physical models outperformed source-trained models for EWT and LMA, and in specific datasets they occasionally exceeded even the best data-driven performance. A model selection framework was developed with 97% accuracy for recommending optimal methods, and Bayesian model averaging (BMA) decision-level fusion further improved estimation accuracy for CHL, CAR, and LMA. The study also demonstrated that a small number of manual measurements (the “target data”) are sufficient to achieve high precision via transfer learning, dramatically reducing the need for extensive field sampling.
This study utilized 26 leaf optical properties datasets from the Ecological Spectral Information System (EcoSIS), spanning North America, Central America, South America, Western Europe, East Asia, and Australia. These datasets collectively included over 30,000 trait–spectral combinations and represented more than 500 species across trees, shrubs, herbs, and vines. Leaf spectra were collected using ASD FieldSpec 3, SVC HR-1024i, and Spectral Evolution PSR+3500 instruments, then standardized to the 400–2400 nm range with 1 nm resolution and smoothed via Savitzky–Golay filter. Five functional traits were targeted: CHL, CAR, EWT, N, and LMA. A leave-one-dataset-out (LODO) validation strategy was employed to ensure robust cross‑dataset evaluation. Four modeling strategies were implemented for data-driven models: small‑sample training (SST), source‑only training (SOT), fine‑tuning transfer learning (STFT), and combined source‑target training (CSTT). Hybrid models replaced source data with 20,000 PROSPECT‑simulated points. Decision‑level fusion of physical and data‑driven outputs was explored using four fusion algorithms, a novel approach not previously reported for leaf trait estimation.
“Our study demonstrates that transfer learning can achieve high‑accuracy leaf trait estimation with surprisingly few field measurements,” the research team led by Shuaipeng Fei and Yuntao Ma explains. “While physical models remain valuable for traits like water content, fine‑tuning data‑driven models on small target datasets offers a practical, resource‑efficient pathway. The 97% accurate model selection framework we developed equips users with clear guidance tailored to their specific data and research objectives.”
The research team sourced 26 global leaf reflectance datasets from the EcoSIS repository, covering diverse ecosystems and instruments. Physical modeling employed PROSPECT‑D and PROSPECT‑PRO inversion. Data‑driven models utilized TabNet, deep residual networks (ResNet), and generalized linear models (GLM) across four training strategies. Hybrid models replaced source data with 20,000 PROSPECT‑simulated radiative transfer points. Decision‑level fusion was performed using four fusion algorithms. A leave‑one‑dataset‑out cross‑validation scheme and Bayesian model averaging (BMA) were applied for robust performance evaluation and ensemble enhancement.
This flexible modeling framework will accelerate the adoption of remote sensing technologies in precision agriculture, enabling farmers to monitor crop nutritional status and water stress at reduced cost and without extensive field campaigns. For global carbon cycle modeling and biodiversity monitoring, the guidelines offer a data‑efficient pathway to upscale functional trait mapping across satellite missions. The demonstrated success of transfer learning from simulated PROSPECT data opens the door to fully synthetic training strategies, ultimately democratizing high‑quality plant trait estimation for resource‑limited researchers worldwide.
###
References
DOI
10.34133/remotesensing.1050
Original Source URL
https://doi.org/10.34133/remotesensing.1050
Funding Information
This work was funded by the Inner Mongolia Grassland Technology Innovation Center Major Innovation Platform Construction Project (CCPTZX2023K03), Key Project of Inner Mongolia Science and Technology Promotion Action (no. NMKJXM202303), and Industrial Technology Innovation Program of IMAST (no. 2024RCYJ04004).
About Journal of Remote Sensing
The Journal of Remote Sensing, an online-only Open Access journal published in association with AIR-CAS, promotes the theory, science, and technology of remote sensing, as well as interdisciplinary research within earth and information science.