Source-linked AI summary
Quantifying vegetation biophysical variables from imaging spectroscopy data: a review on retrieval methods
Jochem Verrelst, Zbyněk Malenovský, Christiaan Van der Tol, Gustau Camps-Valls, Jean-Philippe Gastellu-Etchegorry, Philip Lewis, Peter North, José Moreno
TL;DR
The review addresses how to retrieve vegetation biophysical variables from the large imaging spectroscopy data streams expected from forthcoming satellite missions. It categorizes state-of-the-art methods into four groups and examines multicollinearity, uncertainty, robustness, and processing speed. The review synthesizes these methods and provides recommendations for operational spectroscopy-based processing chains.
Problem
Forthcoming imaging spectroscopy missions will generate large data streams requiring reliable methods for spatiotemporally explicit quantification of vegetation biophysical variables.
Method
The review categorizes imaging spectroscopy retrieval methods as parametric regression, non-parametric regression, physically-based RTM inversion, and hybrid methods.
Results
The review synthesizes the state of the art and identifies multicollinearity, retrieval uncertainties, robustness, and processing speed as central considerations for operational processing.
Takeaways & Limitations
Operational retrieval methods should provide per-pixel uncertainties, while regression methods offer substantially faster full-image processing than physically-based methods.
Takeaways & Limitations
Parametric methods lack associated uncertainty intervals, making retrieval quality difficult to assess across complete images.
Abstract
from arXiv · showhide
An unprecedented spectroscopic data stream will soon become available with forthcoming Earth-observing satellite missions equipped with imaging spectroradiometers. This data stream will open up a vast array of opportunities to quantify a diversity of biochemical and structural vegetation properties. The processing requirements for such large data streams require reliable retrieval techniques enabling the spatiotemporally explicit quantification of biophysical variables. With the aim of preparing for this new era of Earth observation, this review summarizes the state-of-the-art retrieval methods that have been applied in experimental imaging spectroscopy studies inferring all kinds of vegetation biophysical variables. Identified retrieval methods are categorized into: (1) parametric regression, including vegetation indices, shape indices, and spectral transformations; (2) non-parametric regression, including linear and non-linear machine learning regression algorithms; (3) physically-based, including inversion of radiative transfer models (RTMs) using numerical optimization and look-up-table approaches; and (4) hybrid regression methods, which combine RTM simulations with machine learning regression methods. For each of these categories, an overview of widely applied methods with application to mapping vegetation properties is given. In view of processing imaging spectroscopy data, a critical aspect involves the challenge of dealing with spectral multicollinearity. The ability to provide robust estimates, retrieval uncertainties, and acceptable retrieval processing speed are other important aspects in view of operational processing. Recommendations towards new-generation spectroscopy-based processing chains for the operational production of biophysical variables are given.
1 Introduction
Forthcoming imaging spectrometer missions will provide large, spectrally dense data streams for mapping vegetation biophysical variables. The review organizes retrieval methods into four categories while highlighting multicollinearity and operational-processing requirements.
- Quantifying vegetation variables from spectral data requires a model that translates spectral observations into surface biophysical variables.
- Forthcoming EnMAP, HyspIRI, PRISMA, and FLEX missions will produce large spectroscopic data streams for land monitoring and vegetation-variable retrieval.
- Retrieval methods are classified as parametric, non-parametric, physically-based model inversion, and hybrid regression methods.Parametric methods use explicit relationships; non-parametric methods are data-driven; physically-based methods use photon-interaction laws; hybrid methods combine statistical and physically-based elements.
- The categories provide a framework for organizing retrieval methods and imaging spectroscopy applications, although their boundaries are not always clearly defined.Spectral indices, for example, can also serve as inputs to non-parametric methods.
- Spectrally dense data create a key challenge because many retrieval methods must handle spectral multicollinearity, or band redundancy.The review therefore addresses data-processing challenges and possible solutions.
2 Parametric regression methods
Parametric regression links selected spectral measurements or transformations to vegetation variables through explicit fitting functions. Imaging spectroscopy extends these methods with shape and transformation-based approaches, but band reduction, fitting-function choice, and absent uncertainty estimates remain important constraints.
- Discrete spectral band approaches: vegetation indices: Parametric regression explicitly relates a limited number of spectral bands to a biophysical variable through parameterized expressions.Vegetation indices are the oldest and largest group, with intrinsic simplicity as a main advantage.
- Discrete spectral band approaches: vegetation indices: A two-band normalized difference index and linear regression produced an LAI map with R2 of 0.89, RMSE: 0.63, and NRMSE 10.1% in 0.2 seconds.The illustrative HyMap example provided no uncertainty estimates.
- Discrete spectral band approaches: vegetation indices: Reducing full-spectrum data to a few-band index leaves spectral information unexploited and makes optimal band selection and formulation difficult to establish.These concerns arise because discrete-band indices may not capture the complexity of real-world observation conditions.
- Parametric approaches based on spectral shapes and spectral transformations: Shape and transformation methods include red-edge position, derivative-based indices, integration-based indices, continuum removal, and wavelet transforms.These methods extract or enhance information from spectral shapes, regions, absorption features, or multiscale representations.
- Parametric approaches based on spectral shapes and spectral transformations: Red-edge position is the wavelength of maximum first-derivative reflectance between 670 and 780 nm and is sensitive to chlorophyll and structural-variable variation.REP-related methods are typically used to derive canopy chlorophyll content, the product of LAI and leaf chlorophyll content.
- Parametric approaches based on spectral shapes and spectral transformations: Derivative-based indices were concluded to be not necessarily better than conventional and properly elaborated indices.
- Parametric approaches based on spectral shapes and spectral transformations: Spectral transformations still require a fitting function to estimate a biophysical variable, while parametric methods generally provide no associated uncertainty intervals.Their simple formulations support fast processing, but the absence of per-pixel uncertainty makes image-wide retrieval quality difficult to judge.
3 Non-parametric regression methods
Non-parametric regression learns relationships between spectral data and vegetation variables from training data rather than imposing explicit parameterized forms. The review covers linear and nonlinear methods, with reported performance gains often depending on dimensionality reduction or band selection to manage spectral multicollinearity.
- Non-parametric models learn coefficients from training data, avoiding explicit parametrization but potentially requiring more expertise to understand and execute.
- Linear non-parametric methods: Linear methods are computationally fast, but limited samples relative to spectral dimensionality can make covariance estimation and multicollinearity problematic.
- Linear non-parametric methods: PLSR repeatedly outperformed vegetation indices across estimates of biomass, LAI, pigments, fluxes, disease, and nutrient concentrations.
- Nonlinear non-parametric methods: Nonlinear machine-learning regressors capture feature relationships without assuming a particular probability distribution.
- Artificial neural networks: Artificial neural networks outperformed parametric models and, in cited studies, other linear non-parametric models for foliage nitrogen and LAI estimation.
- Kernel-based machine learning regression methods: GPR outperformed most tested machine-learning methods for leaf chlorophyll and LAI, while band selection or dimensionality reduction further improved results.
4 Physically-based model inversion methods
Physically based inversion infers vegetation variables from radiative transfer models using numerical optimization or look-up tables, with physical consistency and uncertainty information but substantial computational cost.
- RTM foundations: Radiative transfer models encode absorption and multiple-scattering physics to infer canopy variables from measured spectra.
- Inversion approaches: Numerical optimization minimizes a cost function by iteratively reducing differences between measured and estimated spectra and variables.Its computational demand increases with RTM complexity and the number of image pixels.
- Inversion approaches: Look-up-table inversion searches simulated RTM realizations, with prior knowledge, cost-function design, and multiple best solutions used to constrain retrievals.Artificial noise can represent uncertainty linked to measurements and models.
- Preprocessing: Spectral selection, smoothing, wavelet transforms, and spectral polishing can improve inversion relationships or resemblance between observed and simulated spectra.
- Strengths and limitations: RTM inversion can improve robustness to solar and view-angle effects and provide uncertainty estimates, but remains slower than statistical methods.LUT approaches may be faster than numerical inversion yet still require computationally expensive per-pixel processing.
5 Hybrid regression methods
Hybrid regression combines RTM simulations with machine-learning regressors to retain physical-model generalization while improving computational efficiency. Results show strong performance for some configurations, while dimensionality reduction and model choice remain important.
- Principles: Hybrid methods train machine-learning regressors on RTM-generated simulations instead of ground data, combining physical generalization with computational efficiency.They still inherit the limitations of the knowledge and concepts represented in the RTM.
- Illustrative results: A PROSAIL-GPR hybrid achieved R2 0.88, RMSE 0.70, and NRMSE 10.1% for LAI mapping with 15% white noise.
- Illustrative results: 6.3 seconds produced a hybrid map using ARTMO’s MLRA toolbox, and GPR also provided uncertainty estimates.Without bare-soil spectra in training, LAI was overestimated over non-irrigated parcels.
- Model choice: Prediction accuracies decreased from numerical optimization to LUT inversion to ANN in one PROSAIL comparison.The comparison did not include dimensionality reduction methods.
- Dimensionality reduction and extensions: PCA improved accuracies over using all bands, while canonical correlation analysis and orthonormalized PLS produced larger improvements.Hybrid structures also extended to PROSPECT-DART, PROSAIL-SVR, and continuum-removal transformations.
6 Discussion
The review organizes imaging-spectroscopy retrieval into four methodological categories and evaluates their suitability for operational processing. Method choice affects retrievability, processing speed, training-data requirements, physical interpretability, uncertainty estimation, and multicollinearity handling.
- Four retrieval categories are summarized: parametric regression, non-parametric regression, radiative-transfer-model inversion, and hybrid methods.Parametric and non-parametric approaches are statistical methods, whereas RTM inversion and hybrid methods rely on RTM simulations.
- Parametric and non-parametric methods require ground data, while RTM inversion uses modeled radiative-transfer knowledge instead of ground measurements.Statistical methods can provide a direct route to target variables when suitable training data are available.
- RTM inversion is most useful when the retrieval targets underlying radiative-transfer processes rather than merely extracting a specific vegetation variable.Examples include scattering, sun- and shade-foliage fractions, canopy light distribution, and relationships between canopy structure and photosynthesis.
- Statistical approaches can relate reflectance to diverse measured biochemical and structural variables, but model validity and transferability typically depend on correlations with validation data.This flexibility lacks a physical basis, which motivates caution when relying solely on best correlations.
- Bayesian statistical models can provide prediction uncertainties, and dimensionality reduction can accelerate LAI mapping while lowering per-pixel uncertainties.Uncertainty maps can also help evaluate model transferability across space and time.
- RTM inversion is computationally expensive because simulations and iterative per-pixel fitting require repeated processing; emulators improve simulation speed but remain slower than statistical methods.Initial emulator experiments preserved sufficient accuracy relative to original RTMs, but per-pixel spectral fitting still requires many repetitions.
- Operational processing requires per-pixel uncertainties, fast mapping, and strategies such as band selection or dimensionality reduction to mitigate regression multicollinearity.Statistical methods can process full images in minutes or seconds, while physically based methods avoid spectral multicollinearity but are slower.
7 Conclusions
The review synthesizes four retrieval-method categories for mapping vegetation properties from imaging spectroscopy and assesses their operational suitability. It highlights uncertainty-aware nonlinear regression, while identifying uncertainty estimation and multicollinearity as central processing considerations.
- The review synthesizes parametric regression, non-parametric regression, physically based RTM inversion, and hybrid methods applied to imaging-spectroscopy vegetation mapping.
- Parametric methods extract spectroscopic information effectively but lack uncertainty estimates, limiting their suitability for operational use.
- Nonlinear non-parametric methods, especially probabilistic machine-learning approaches such as Gaussian process regression, can reach higher accuracies.
- Spectral multicollinearity remains a central challenge for regression and hybrid methods, whereas physically based spectral-fitting methods do not suffer from it.