Tagliabue et al. (2025) Appraising retrieval schemes from spaceborne hyperspectral imagery for mapping leaf and canopy traits in forest ecosystems

Identification

• Journal: Remote Sensing of Environment
• Year: 2025
• Date: 2025-11-27
• Authors: Giulia Tagliabue, Cinzia Panigada, Beatrice Savinelli, Luigi Vignali, Micol Rossini
• DOI: 10.1016/j.rse.2025.115145

Research Groups

• Department of Earth and Environmental Sciences, University of Milano - Bicocca, Milan, Italy
• National Biodiversity Future Center (NBFC), Palermo, Italy

Short Summary

This study investigated and compared machine learning regression algorithms (MLRA) and hybrid approaches for retrieving forest traits from PRISMA hyperspectral imagery. It demonstrated that hybrid models accurately quantify key leaf and canopy traits, including Leaf Chlorophyll Content (LCC), Leaf Nitrogen Content (LNC), Leaf Water Content (LWC), Leaf Mass per Area (LMA), and Leaf Area Index (LAI), in forest ecosystems, even under drought conditions.

Objective

• To develop, test, and compare retrieval schemes for forest traits (Leaf Chlorophyll Content (LCC), Leaf Carotenoid Content (Ccx), Leaf Nitrogen Content (LNC), Leaf Water Content (LWC), Leaf Mass per Area (LMA), and Leaf Area Index (LAI)) in forest ecosystems using state-of-the-art machine learning regression algorithms (MLRA) and hybrid approaches combining MLRA with radiative transfer simulations.
• To evaluate the ability of these models to detect seasonal variations in plant traits driven by environmental stressors, such as the 2022 drought.

Study Configuration

• Spatial Scale: Ticino Regional Park, northern Italy, covering approximately 918 square kilometers of temperate mixed forest. Field data were collected from 50 stands, each measuring 30 meters × 30 meters. PRISMA imagery has a nominal ground sampling distance (GSD) of 30 meters.
• Temporal Scale: Summer 2022 (June to September). Four PRISMA overpasses for the northern part (June 11, July 27, August 31, September 6, 2022) and three for the central part (July 10, August 2, September 11, 2022). Intensive field campaigns were conducted concurrently with these overpasses.

Methodology and Data

• Models used:
  • Machine Learning Regression Algorithms (MLRA): Gaussian Processes Regression (GPR), Kernel Ridge Regression (KRR), Neural Networks (NN), Principal Component Regression (PCR), Partial Least Squares Regression (PLSR), and Support Vector Regression (SVR).
  • Radiative Transfer Models (RTMs): PROSPECT-PRO (leaf level) coupled with INFORM (canopy level, a hybrid turbid medium-geometric model combining SAIL and FLIM).
  • Hybrid Models: Combined PROSPECT-PRO-INFORM simulations with GPR, KRR, and SVR.
• Data sources:
  • Spaceborne Hyperspectral Imagery: PRISMA satellite data (400 nm to 2500 nm, 240 spectral channels, ≤12 nm spectral resolution, 30 m GSD).
  • Field Data: Ground measurements from 50 forest stands collected in summer 2022. Traits measured include:
    • Leaf Chlorophyll Content (LCC) (in g m⁻²)
    • Leaf Carotenoid Content (Ccx) (in g m⁻²)
    • Leaf Nitrogen Content (LNC) (in g m⁻²)
    • Leaf Water Content (LWC) (in kg m⁻²)
    • Leaf Mass per Area (LMA) (in kg m⁻²)
    • Leaf Area Index (LAI) (in m² m⁻², unitless) LAI was quantified using digital hemispherical photos processed with CAN-EYE.
  • Synthetic Data: A Look-Up Table (LUT) of 2000 synthetic spectra generated by running the coupled PROSPECT-PRO-INFORM model in forward mode.

Main Results

• Hybrid models generally showed slightly superior performance compared to purely statistical MLRA approaches, particularly for Leaf Chlorophyll Content (LCC) and Leaf Nitrogen Content (LNC).
• Kernel-based algorithms (GPR, KRR, SVR) consistently outperformed other MLRA (NN, PCR, PLSR) in both pure MLRA and hybrid schemes.
• Hybrid models accurately quantified most forest traits:
  • Leaf Chlorophyll Content (LCC): r² = 0.67, nRMSE = 13.5 %
  • Leaf Nitrogen Content (LNC): r² = 0.82, nRMSE = 9.5 %
  • Leaf Water Content (LWC): r² = 0.98, nRMSE = 3.5 %
  • Leaf Mass per Area (LMA): r² = 0.93, nRMSE = 6.6 %
  • Leaf Area Index (LAI): r² = 0.83, nRMSE = 11.7 %
• Leaf Carotenoid Content (Ccx) yielded lower accuracy (r² = 0.43, nRMSE = 13.5 %), indicating challenges in decoupling its spectral contribution from LCC.
• The models successfully captured drought-induced temporal variations: a significant decrease in LCC, Ccx, and LAI was observed in early September 2022 due to the severe drought, while LNC, LWC, and LMA remained relatively stable, consistent with field observations.
• Spectral sensitivity analysis of GPR-based hybrid models revealed trait-specific dependencies:
  • LCC: Dominant contribution from the visible (VIS) spectral region (400–750 nm).
  • LNC: Substantial contributions from both the shortwave infrared (SWIR) (1300–2500 nm) and VIS regions.
  • LMA and LWC: Strong dependence on the near-infrared (NIR) (750–1300 nm) and SWIR regions.
  • LAI: Almost equal contribution from VIS, NIR, and SWIR spectral domains.

Contributions

• This is the first study to evaluate hybrid retrieval schemes using real spaceborne hyperspectral data (PRISMA) for mapping a comprehensive set of multiple leaf and canopy traits in forest ecosystems.
• It provides a valuable reference framework for the exploitation of forthcoming global hyperspectral missions (e.g., ESA's CHIME, NASA's SBG) for operational forest monitoring.
• Demonstrated the effectiveness of hybrid models in operational conditions, including their ability to detect trait-specific variability induced by environmental stressors like drought, even when traits become uncorrelated (e.g., LCC and LNC).
• Highlighted the superior performance and robustness of kernel-based machine learning algorithms within hybrid frameworks compared to purely statistical approaches, especially when dealing with potentially unbalanced or unrepresentative training data.

Funding

• PRIS4VEG project of the Italian Space Agency (ASI Contract No. 2022–5-U. 0)
• National Biodiversity Future Center-NBFC project (Project Code No. CN-00000033, CUP: H43C22000530001) funded by European Union-NextGenerationEU

Citation

@article{Tagliabue2025Appraising,
  author = {Tagliabue, Giulia and Panigada, Cinzia and Savinelli, Beatrice and Vignali, Luigi and Rossini, Micol},
  title = {Appraising retrieval schemes from spaceborne hyperspectral imagery for mapping leaf and canopy traits in forest ecosystems},
  journal = {Remote Sensing of Environment},
  year = {2025},
  doi = {10.1016/j.rse.2025.115145},
  url = {https://doi.org/10.1016/j.rse.2025.115145}
}