paper/paper.md

In jbferet/prospect: PROSPECT leaf radiative transfer model and inversion routines

title: 'prospect: an R package to link leaf optical properties with their chemical and structural properties with the leaf model PROSPECT' tags: - R - leaf spectroscopy - vegetation biophysical properties - remote sensing - physical modeling - iterative optimization authors: - name: Jean-Baptiste Féret orcid: 0000-0002-0151-1334 corresponding: true affiliation: 1 - name: Florian de Boissieu orcid: 0000-0002-2185-9952 affiliation: 1 affiliations: - name: TETIS, INRAE, AgroParisTech, CIRAD, CNRS, Université Montpellier, Montpellier, France index: 1 date: 17 July 2023 bibliography: paper.bib

Summary

PROSPECT simulates optical properties for a leaf, based on its content in absorbing constituents, and a structure parameter accounting for light scattering. We present prospect, an R package aiming at providing the most recent versions of the model, along with inversion routines to assess leaf chemistry from optical properties.

Statement of need

The capacity to measure, map and monitor vegetation traits corresponding to biophysical and chemical properties is crucial to better understand ecosystem and agrosystem functions, as well as carbon, water and energy budgets. At leaf scale, these vegetation traits are linked to their optical properties through absorption and scattering mechanisms. Spectroscopic techniques provide rapid, accurate and non-destructive assessment of leaf chemical composition. Physical models aim at simulating optical properties of leaves from their chemical composition. Inversion techniques take advantage of these models to assess leaf chemical composition from optical properties. The model PROSPECT (leaf optical PROperties SPECtra) is currently the most popular physical model for the simulation of leaf optical properties. PROSPECT simulates leaf directional-hemispherical reflectance and transmittance from a combination of chemical constituents and their corresponding specific absorption coefficients. It uses a simplified representation of leaf structure through the generalised plate model [@allen1970] to account for scattering. PROSPECT can be coupled with canopy reflectance models to analyze Earth observation data. Hence, it is a key component for remote sensing applications dedicated to vegetation monitoring.

Multiple versions have been released since its first version [@jacquemoud1990]. @feret2008 introduced carotenoids and @feret2017 introduced anthocyanins, to simulate leaf optical properties from juvenile to mature and senescent development stages. @feret2021 introduced PROSPECT-PRO, the latest version separating dry matter constituents into proteins and carbon based constituents. In parallel with updated versions of the model, model inversion strategies have been introduced to improve the assessment of leaf chemical constituents [@feret2019; @spafford2021].

PROSPECT implementations since @feret2008 can be found at this webpage. This includes distributions in matlab, R and fortran programming languages. PROSPECT is coupled with vegetation models, including COSINE dedicated to close-range imaging spectroscopy [@jay_physically-based_2016] and SAIL for canopy reflectance modeling [verhoef_coupled_2007, @verhoef_unified_2007, @jacquemoud_prospect+_2009]. A version of prospect coupled with SAIL is available in the R package prosail. Note that PROSPECT and PROSAIL are also available in packages written in python, Julia and R.

Overview

PROSPECT simulation in forward mode

In forward mode, PROSPECT simulates leaf directional-hemispherical reflectance and transmittance from a set of chemical constituents and a leaf structure parameter identified as N. N represents the number of homogeneous layers introduced in the generalized plate model.

Two versions of PROSPECT are implemented in prospect: PROSPECT-D [@feret2017] and PROSPECT-PRO [@feret2021]. PROSPECT-4 and 5 are deprecated. Table \ref{table:1} lists the input leaf chemical constituents and structure parameter for versions D and PRO available with the function PROSPECT.

| Version | D | PRO | |--------- |:----: |:----: | | CHL | X | X | | CAR | X | X | | ANT | X | X | | BROWN | X | X | | EWT | X | X | | LMA | X | | | PROT | | X | | CBC | | X | | N | X | X |

: Input variables for the two versions of PROSPECT implemented in the prospect package (CHL: chlorophylls; CAR: carotenoids; ANT: anthocyanins; BROWN: brown pigments; EWT: equivalent water thickness; LMA: leaf mass per area; PROT: proteins; CBC: carbon based constituents; N: leaf structure parameter).\label{table:1}

PROSPECT is a relatively simple and computationally efficient model. Hence iterative optimization can be considered to invert PROSPECT and assess leaf chemistry and structure from their optical properties. Iterative optimization aims at minimizing a merit function comparing measured and simulated leaf optical properties. This procedure is based on the function fmincon included in the package pracma. It usually takes less than 1 second to perform PROSPECT inversion.

Various inversion strategies using iterative optimization are described in the literature [@colombo_estimation_2008, @li_retrieval_2011, @feret2019]. These inversion strategies differ either by the merit function, or by the selection of specific spectral domains used to retrieve one or several leaf biophysical properties, or by the introduction of prior information. The default merit function, Merit_PROSPECT_RMSE, corresponds to the root mean square of the mean quadratic difference between measured and simulated leaf optical properties (reflectance and/or transmittance). Alternative merit functions using the same input and output variables as Merit_PROSPECT_RMSE can be defined. The online documentation provides an example of alternative merit function.

Table \ref{table:2} details the optimal spectral domains to assess leaf chemical constituents from their optical properties [@feret2019; @spafford2021].

| Constituent| Optimal spectral domain | Versions | |----------- |--------------------------------- |:-----------: | | CHL | 700 -- 720 | D, PRO | | CAR | 520 -- 560 | D, PRO | | ANT | 400 -- 800 | D, PRO | | BROWN | NA | NA | | EWT | 1700 -- 2400 | D, PRO | | LMA | 1700 -- 2400 | D | | PROT | 2100 -- 2139; 2160 -- 2179 | PRO | | CBC | 1480 -- 1499; 1560 -- 1579; 1760 -- 1799; 2040 -- 2059; 2120 -- 2139; 2160 -- 2239; 2260 -- 2279; 2340 -- 2359; 2380 -- 2399 | PRO |

: Optimal spectral domains to assess vegetation chemical constituents from leaf optical properties.\label{table:2}

Example 1: running PROSPECT in forward mode

Input parameters listed in Table \ref{table:1} are required to simulate leaf optical properties with PROSPECT. Optical constants corresponding to specific absorption coefficients and leaf refractive index are identical for all leaves, and are defined over the spectral domain ranging from 400 nm to 2500 nm. They are accessible through the variable SpecPROSPECT_FullRange, a data frame automatically loaded with the package. Multiple functions in the package prospect expect a variable SpecPROSPECT as input. SpecPROSPECT_FullRange is used as default value for this input variable. Simulated spectral characteristics, including spectral domain, sampling and sensor-specific spectral response can be derived from SpecPROSPECT_FullRange.

The function PROSPECT identifies the version to be used: PROSPECT-D is used if LMA is defined, while PROSPECT-PRO is used if proteins (PROT) and carbon based constituents (CBC) are defined. If LMA, PROT and CBC are defined simultaneously, PROSPECT-PRO is used and LMA is set to 0. Figure \ref{fig:LOP} compares simulated leaf optical properties. Here, the differences between PROSPECT-D and PROSPECT-PRO are mainly driven by the different value set for the N structure parameter.

# Load prospect package
library(prospect)
# Run PROSPECT-D
LRT_D <- PROSPECT(CHL = 45, CAR = 10, ANT = 0.2, 
                  EWT = 0.012, LMA = 0.010, N = 1.3)
# Run PROSPECT-PRO
LRT_PRO <- PROSPECT(CHL = 45, CAR = 10, ANT = 0.2, 
                    EWT = 0.012, PROT = 0.001, CBC = 0.009, N = 1.7)

{ width=85% }

The spectral domain covered with PROSPECT simulations can be adjusted with the function FitSpectralData, which adjust information from SpecPROSPECT_FullRange to a user-defined spectral domain.

# define spectral bands for Visible / Near InfraRed (VNIR) simulation
wvlRange_VNIR <- seq(400,1000)
# adjust spectral properties used in PROSPECT to VNIR domain
VNIR <- FitSpectralData(lambda = wvlRange_VNIR)
# Run PROSPECT-D in VNIR domain
LRT_VNIR <- PROSPECT(SpecPROSPECT = VNIR$SpecPROSPECT,
                     N = 1.4, CHL = 30, CAR = 6, EWT = 0.02, LMA = 0.01)

Example 2: PROSPECT inversion using iterative optimization

Inversion is illustated with the experimental database ANGERS, a publicly available dataset used to calibrate PROSPECT. ANGERS can be downloaded from a gitlab repository with the function download_LeafDB.

LeafDB <- download_LeafDB(dbName = 'ANGERS')

PROSPECT inversion using the full spectral information

The spectral domains covered by SpecPROSPECT and the leaf optical properties are expected to match when performing inversion over the full spectral domain covered by the leaf optical properties. The function FitSpectralData harmonizes the spectral domain for SpecPROSPECT and the leaf optical properties. A vector listing the wavelengths of interest can also be provided with UserDomain. The following R code adjusts SpecPROSPECT to the spectral domain of the leaf optical properties.

# Adjust spectral domain for SpecPROSPECT to fit leaf optical properties
SubData <- FitSpectralData(SpecPROSPECT = SpecPROSPECT_FullRange,
                           Refl = LeafDB$Refl, 
                           Tran = LeafDB$Tran,
                           lambda = LeafDB$lambda, 
                           UserDomain = LeafDB$lambda)

The main inversion procedure is called with the function Invert_PROSPECT, which minimizes a merit function defined by Merit_PROSPECT_RMSE as default option. Part or all parameters from Table \ref{table:1} can be assessed. Parameters assessed from PROSPECT inversion are defined with an input list, Parms2Estimate. The default parameterization of PROSPECT inversion leads to the assessment of all parameters listed in Table \ref{table:1} except BROWN. BROWN can be assessed during inversion (if working with senescent leaves) by setting Est_Brown_Pigments = TRUE as input for Invert_PROSPECT. The value of unassessed parameters is set in InitValues.

# Assess all parameters using PROSPECT inversion applied to full spectral data
res_all_WL <- Invert_PROSPECT(SpecPROSPECT = SubData$SpecPROSPECT, 
                              Refl = SubData$Refl,  
                              Tran = SubData$Tran)

PROSPECT inversion using optimal spectral domains for each constituent

The function Invert_PROSPECT_OPT performs PROSPECT inversion using optimal spectral domains, specific to each constituent [@feret2019; @spafford2021]. The function automatically adjusts the spectral domain of the leaf optical properties provided as input.

# Assess a set of parameters using PROSPECT inversion with optimal spectral domains
Parms2Estimate <- c('CHL', 'CAR', 'EWT', 'LMA')
res_opt_WL <- Invert_PROSPECT_OPT(lambda = LeafDB$lambda, 
                                  Refl = LeafDB$Refl, 
                                  Tran = LeafDB$Tran
                                  Parms2Estimate = Parms2Estimate)

Performances of the two types of inversion: Comparison with ANGERS data

Figure \ref{fig:scatter} compares lab measurements from CHL, CAR, EWT and LMA with values assessed from Invert_PROSPECT and Invert_PROSPECT_OPT using ANGERS.

{ width=90% }

Conclusion

We introduce prospect, an R package dedicated to the PROSPECT leaf model. prospect runs various versions of the model in direct mode to simulate directional-hemispherical reflectance and transmittance. It also includes inversion routines to assess leaf structure and chemical constituent content from reflectance and/or transmittance. prospect provides latest advances in terms of model version and inversion procedures to the leaf spectroscopy community. prospect is coupled with the canopy model SAIL through the R package prosail. prosail is dedicated to applications focusing on Earth observation imagery analysis and allows simulation of canopy reflectance for multispectral and hyperspectral sensors. Hybrid inversions based on physical modeling and machine learning are also implemented in prosail to assess vegetation traits at canopy scale from imagery data.

Availability

prospect is an open-source software package made available under the MIT license. Tutorials are available at https://jbferet.gitlab.io/prospect/.

Acknowledgements

The authors acknowledge financial support from Agence Nationale de la Recherche (BioCop project — ANR-17-CE32-0001). We are grateful to Stéphane Jacquemoud and Frédéric Baret for the development of the initial version of the PROSPECT model. We warmly thank Luc Bidel, Christophe François and Gabriel Pavan who collected the ANGERS data set. We warmly thank Christopher Anderson, Alexey Shiklomanov and Neha K. Nawandar for their review and the constructive comments and suggestions they provided, which greatly helped in improving both performances and functionalities of this package.

References

jbferet/prospect documentation built on Feb. 10, 2025, 9:35 a.m.