Les commentaires sont clos.

CDD in remote sensing and physical modelling of mediterranean forests

8 Septembre 2022

Catégorie : Ingénieur

Title : "Estimation of Mediterranean forest functional traits : a multi-sensor assessment in function of the canopy cover"

Associated project: APR CNES TOSCA SentHyMED "Synergy between Sentinel-2 multi-temporal imagery and Hyperspectral imagers for a better monitoring of functional traits of Mediterranean forests" (ONERA, CESBIO, TETIS, DYNAFOR, CEFE)(2021-2024)

Duration : 12 months (possibility to renew for 12 months more in the framework of the project)


Spectroscopy, Sentinel-2, LiDAR data, 3D modelling, radiative transfer, inversion, Mediterranean forests méditerranéennes

Formation and skills

Engineering schools and/or research masters: signal/image or remote sensing data processing, machine learning methods, ecology/environment, python programming

Notions in radiative transfer modeling, 3D LiDAR information processing

A first experience on the use of the DART code (but not a prerequisite)


Mediterranean forests cover a relatively small area on a global scale for a high concentration of plant diversity, including endemic species that have been able to adapt to the hazards of the Mediterranean climate (long, hot and dry summers, mild and wet winters). However, they reprensent critical areas, because by 2100 their biodiversity will change much more than that of other terrestrial ecosystems [1], in particular due to the increase in fires, droughts, and anthropogenic exploitation (urbanization, wood industry, increase in agricultural lands) [2]. It is therefore important to monitor the temporal evolution of these forests and the health of some of their tree species, in order to guide environmental policies and preserve the biodiversity of these fragile environments. Remote sensing observations can be used to address this issue, as well as for large-scale mapping of essential biodiversity variables (EBVs) defined by groups of scientific experts (GEO-BON)[3] and, in particular, the study of the "EBV-species traits" class, which aims to characterize the functional traits of species through their phenological, morphological and physiological attributes. The monitoring of these traits is partly achieved through the estimation of certain biophysical-chemical properties at the leaf and canopy scale accessible by remote sensing. Among these properties, the most commonly studied are leaf pigments (CAB - chlorophylls and CAR - carotenoids), leaf water and dry matter content (EWT, LMA), and leaf area index (LAI). Over a phenological cycle of the plant, the latter are indicators that allow the evaluation of water stress, the observation of biomass changes and early defoliation.

With the multiplication of remote sensing data from current and future satellites, a major challenge is to assess the potential of combining all these datasets to better describe and study vegetation, given the spectral, spatial and temporal characteristics of the different sensors [4]. For our case study, the simultaneous estimation of the above-mentioned biophysical-chemical properties requires a high spectral richness to take advantage of their sensitivity over the whole measured spectrum. With its larger number of spectral bands, hyperspectral imaging over [0.4-2.5 µm] is more suitable than multispectral imaging, which however often has a shorter revisit time. Many studies have already evaluated the performance of these properties for dense forests in tropical and temperate climates [5-6], but few have focused on sparse forests or forests with a large variability in tree cover (e.g. Mediterranean climate forests). Indeed, for the latter case, a spatial resolution that is not fine enough (decametric order) may reveal mixed pixels whose reflectance is due to the contribution of several elements (i.e. tree, soil, understory). One solution to estimate the biophysical-chemical properties of tree vegetation is the use of hybrid inversion methods that combine radiative transfer models (RTM), machine learning methods and remote sensing data. The RTM simulates the remotely sensed images of the forest canopy based on the structural and optical properties of its elements (e.g. tree height and distribution, leaf reflectance and transmittance, etc.). By varying these properties, the RTM can create a database of simulated reflectances that can be used to train and validate a learning method (PLSR, RFR, SVMR, ANN). This method is then applied to spectral images to produce vegetation trait maps that are compared with field measurements for validation purposes.

A current major challenge is to determine the level of complexity required to model a tree scene according to the forest canopy cover (RTM parameterisation), the most appropriate learning method for all the vegetation traits studied, and the spectral and spatial adaptations to be carried out between Sentinel-2 multi-spectral data at 10/20m and hyperspectral missions at 10m (BIODIVERSITY) and 30m (PRISMA, EnMAP, CHIME, SBG)

The proposed work aims to address these issues using the 3D canopy-scale RTM code DART [7], including the leaf-scale RTM code PROSPECT [8] (link between the optical properties of the leaf and the leaf traits). It is part of the second year of the TOSCA CNES SentHyMED APR project [9], which aims to characterize the health status and water stress of two Mediterranean forests located to the north of Montpellier (Puéchabon and Pic Saint Loup) with two main oak species (deciduous: white oak and evergreen: green oak). The work will consider a given date corresponding to the MEDOAK measurement campaign (June 2021), carried out in parallel with NASA-ESA airborne hyperspectral acquisitions for the CHIME and SBG missions.

Initially, the work will focus on hyperspectral data:

1) Simulation of satellite images at 10m (BIODIVERSITY) and 30m (CHIME, SBG) from AVIRIS-Next Generation airborne images at 1 and/or 3m using existing tools at ONERA that allow to consider instrumental noise,

2) Creation of 3D DART mock-ups of the study plots on the two sites, using LiDAR drone data and forest inventories. The LiDAR point clouds will be converted into a 3D matrix of "turbid" voxels with the Amapvox tool. These 3D matrices will then be used to create the DART mock-ups,

3) Generation of simulated spectral databases with DART from the created mock-ups and building of a design of experiments for the variation of scene properties,

4) Training of machine learning methods and their application on the whole set of images (airborne and satellite simulated) for the estimation and mapping of biophysical-chemical properties. Methods are already available through work on Mediterranean forest sites in California (Thomas Miraglio's thesis [10-12] and APR CNES HyperMED [13]) and tropical forest sites in French Guyana (Dav Ebengo's thesis [14] and APR CNES HyperTropik [15]),

5) Comparison with field data for performance evaluation.

In a second phase, the work will focus on Sentinel-2 data:

1) Selection of images on the THEIA platform acquired as close as possible to the MEDOAK campaign,

2) Co-registration with airborne hyperspectral images using the Gefolki tool,

3) Spectral and spatial resampling of the DART-simulated spectral databases to match the Sentinel-2 data characteristics,

4) Inversion with the above methods but adapted to the spectral indices due to the reduction of the number of spectral bands,

5) Comparison of the results with the field data.


Expected products :

⮚Estimation maps of LAI, CAB, CAR, EWT and LMA

Expected results :

⮚ Review of the impact of spatial resolution and instrumental characteristics on the performance of biophysical-chemical property estimation,

⮚ Review of the best machine learning method for inversion,

⮚ Assessment of the performance of biophysical-chemical properties estimation according to the forest canopy cover.

This 12-month short-term contract will be on the CESBIO payroll and hosted at the ONERA centre in Toulouse. It may be followed by a 12-month another short-term contract at the DYNAFOR laboratory in Toulouse to take into account multi-temporal aspects in the method developed and by adding phenological indicators to assess the water stress of these forests.

If you are interested, please send your CV and motivation letter to J.P. Gastellu-Etchegorry ( and K. Adeline (


[1] O. E. Sala et al., “Global biodiversity scenarios for the year 2100,” Science. 2000, doi: 10.1126/science.287.5459.1770.

[2] E. C. Underwood, J. H. Viers, K. R. Klausmeyer, R. L. Cox, and M. R. Shaw, “Threats and biodiversity in the mediterranean biome,” Divers. Distrib., 2009, doi: 10.1111/j.1472-4642.2008.00518.x.

[3] H. M. Pereira et al., “Essential biodiversity variables,” Science. 2013, doi: 10.1126/science.1229931.

[4] J. Transon, R. d’Andrimont, A. Maugnard, P. Defourny. Survey of hyperspectral Earth Observation applications from space in the Sentinel-2 context. Remote Sens. Vol. 10:2, 1-32, 2018, doi: 10.3390/rs10020157.

[5] M. P. Ferreira, J. B. Féret, E. Grau, J. P. Gastellu-Etchegorry, Y. E. Shimabukuro, and C. R. de Souza Filho, “Retrieving structural and chemical properties of individual tree crowns in a highly diverse tropical forest with 3D radiative transfer modeling and imaging spectroscopy,” Remote Sens. Environ., 2018, doi: 10.1016/j.rse.2018.04.023.

[6] F. D. Schneider et al., “Mapping functional diversity from remotely sensed morphological and physiological forest traits,” Nat. Commun., vol. 8, no. 1, 2017, doi: 10.1038/s41467-017-01530-3.


[8] J.-B. Féret, K. Berger, F. De Boissieu, and Z. Malenovský, “PROSPECT-PRO for estimating content of nitrogen-containing leaf proteins and other carbon-based constituents,” ArXiv200311961 Q-Bio, 2020.

[9] ONERA, CESBIO, TETIS, DYNAFOR and CEFE, « APR TOSCA CNES SentHyMED « Complémentarité entre de l'imagerie multi-temporelle Sentinel-2 et des imageurs Hyperspectraux pour un meilleur suivi des traits fonctionnels de forêts MEDiterranéennes », 2021-2023

[10] T. Miraglio, K. Adeline, M. Huesca, S. Ustin, and X. Briottet, “Monitoring LAI, Chlorophylls, and Carotenoids Content of a Woodland Savanna Using Hyperspectral Imagery and 3D Radiative Transfer Modeling,” Remote Sens., vol. 12, no. 1, p. 28, Dec. 2019, doi: 10.3390/rs12010028.

[11] T. Miraglio et al., “Impact of modeling abstractions when estimating leaf mass per area and equivalent water thickness over sparse forests using a hybrid method,” Remote Sens., vol. 13, no. 16, 2021, doi: 10.3390/rs13163235.

[12] T. Miraglio et al., “Assessing Vegetation Traits Estimates Accuracies from the Future SBG and Biodiversity Hyperspectral Missions Over two Mediterranean Forests”. International Journal of Remote Sensing, 43:10, 3537-3562, doi: 10.1080/01431161.2022.2093143.

[13] ONERA, CESBIO, CSTARS, BiometLab. APR TOSCA CNES HyperMED: Evaluation des caractéristiques fonctionnelles des essences d’arbres pour le suivi de leur état de santé pour des écosystèmes de forêts MEDiterranéennes pour un imageur Hyperspectral. Préparation de la mission hyperspectrale HYPS. 2019-2022.

[14] Dav Ebengo Mwampongo, “Apport de la modélisation physique pour la cartographie de la biodiversité végétale en forêts tropicales par télédétection optique,” Université de Montpellier.

[15] TETIS; CESBIO; AMAP; ECO&SOLS; Orsay, E. APR TOSCA CNES HyperTropik “Estimation de la biodiversité des forêts tropicales par imagerie hyperspectrale”, préparation de la mission hyperspectrale HYPXIM 2014