Résumé
Characterizing the condition of Mediterranean forests is crucial for fire risk prevention, and for assessing the impact of increasing droughts and changes in land use. Optical remote sensing provides access to various biophysical and biochemical vegetation properties, including essential biodiversity variables. This capacity enhances our understanding of the structure and functioning of these forests. However, the accuracy these properties can be estimated depends on several factors: the scale of observation (i.e. the spatial resolution), the spectral richness (i.e. the number of spectral bands, sampling and resolution) and the revisit time (for monitoring purposes) of the remote sensing sensors. Hybrid inversion methods are increasingly used as they require fewer field measurements compared to empirical methods. These methods adopt a more generalizable and physically-based approach by relying on the combination of radiative transfer model (RTM) simulations and machine learning regression methods (MLRA). Key challenges include the adequate choice of the RTM (e.g. 1D, 2D or 3D) and the accurate input parameterization of the RTM, in order to account for the particularities of the observed landscape and to simulate realistic remote sensing images (e.g. differences between processing dense or open forests). The objective of this study is to compare the accuracy of vegetation trait estimations for the tree overstory layer based on a hybrid inversion using either 1D or 3D RTM and a multi-sensor dataset encompassing multi-/hyperspectral and airborne/satellite data. Two Mediterranean forests located in the South of France are studied: Pic Saint Loup and Puéchabon. We focused on two oak species: Quercus ilex (evergreen) and Quercus pubescens (deciduous). The MEDOAK campaign [1] occurred in June 2021 and resulted in the collection of a variety of field and lab measurements, including overstory and understory inventory, plant area index, spectroscopic data, leaf-clip optical data, trait measurements and inversions with the leaf model PROSPECT [2][3]. This ground data collection was carried out simultaneously to the Hypersense campaign, a joint airborne campaign organized by NASA and ESA, and managed by the University of Zurich in cooperation with NASA/JPL to support the upcoming Copernicus Hyperspectral Imaging Mission for the Environment (CHIME) and Surface Biology and Geology (SBG) satellite missions. Airborne imaging spectroscopy data were acquired with AVIRIS-Next Generation sensor at 1m and 3m spatial resolutions. These airborne acquisitions were complemented with Sentinel-2, PRISMA and DESIS acquisitions over the same period. In addition, UAV LiDAR scans were acquired on some forest plots. Past work aimed at designing detailed 3D forest mockups from these LiDAR point clouds, assessing the accuracy of the simulations performed with the DART 3D RTM [4], and studying the contribution of factors including understory reflectance, woody element and abstract modeling of leaf optical properties in total canopy reflectance [5]. The current work focuses on identifying the optimal geometrical parameterization for vegetation trait estimations by comparing: (1) a homogeneous and simple representation with PROSAIL 1D RTM [6], (2) a simplified forest representation using geometric objects with DART and (3) a detailed 3D representation obtained from LiDAR mockups with DART. Targeted vegetation traits are plant area index, leaf pigments, water and dry matter content, and live fuel moisture content. Inversion methods are based on existing tools [7] and based on [8]. For comparison purposes, several MLRAs are compared, including PLSR, SVMR and RFR. The results are explored and discussed between the three different geometrical modelling strategies, the remote sensing sensor characteristics (spatial and spectral), and the studied forest plot characteristics (canopy cover, species composition). This work also aims at preparing the French hyperspectral satellite mission BIODIVERSITY, featuring 10 m spatial resolution [9][10]. Therefore, 10 m images were simulated from 1m airborne data and comparisons were performed to assess the potential of improved spatial resolution compared to current operational hyperspectral missions at 30 m. To conclude, these results will contribute to assess the complementarity of remote sensing data (multi- vs hyperspectral, airborne vs satellite) to map vegetation traits useful to determine forest water stress and risk prevention through their monitoring. [1] K. Adeline, J.-B. Féret, H. Clenet, J.-M. Limousin, J.-M. Ourcival, F. Mouillot, S. Alleaume, A. Jolivot, X. Briottet, L. Bidel, E. Aria, A. Defossez, T. Gaubert, J. Giffard-Carlet, J. Kempf, D. Longepierre, F. Lopez, T. Miraglio, J. Vigouroux and M. Debue (2024). Multi-scale datasets for monitoring Mediterranean oak forests from optical remote sensing during the SENTHYMED/MEDOAK experiment in the north of Montpellier (France). Data in Brief, 53, 110185. [2] Féret J-B, Gitelson AA, Noble SD and Jacquemoud S (2017). PROSPECT-D: Towards modeling leaf optical properties through a complete lifecycle. Remote Sensing of Environment, 193, 204–215. [3] J.-B. Féret, J. Giffard-Carlet, S. Alleaume, X. Briottet, V. Chéret, H. Clénet, J.-P. Denux, J.-P. Gastellu-Etchegorry, A. Jolivot, J.-M. Limousin, F. Mouillot, J.-M. Ourcival and K. Adeline (2022). Estimating functional traits in Mediterranean ecosystems using spectroscopy from leaf to canopy scale. 2nd Workshop on International Cooperation in Spaceborne Imaging Spectroscopy, 19-21 october 2022, Frascati, Italy, oral. [4] Yingjie Wang, Abdelaziz Kallel, Xuebo Yang, Omar Regaieg, Nicolas Lauret, Jordan Guilleux, Eric Chavanon, Jean-Philippe Gastellu-Etchegorry (2022). DART-Lux: An unbiased and rapid Monte Carlo radiative transfer method for simulating remote sensing images. Remote Sensing of Environment, Volume 274, 2022, 112973. [5] M. Debue, G. Vincent, S. Alleaume, F. de Boissieu, X. Briottet, J.-B. Féret, J.-P. Gastellu-Etchegorry, J.-M. Limousin, D. Longepierre and K. Adeline. Adequacy of Mediterranean forest simulations from DART radiative transfer model and UAV laser scanning data to multi- and hyperspectral images. SPIE Remote Sensing, 3-6 September 2023, Amsterdam, Netherlands, oral and proceeding. [6] Jacquemoud S, Verhoef W, Baret F, Bacour C, Zarco-Tejada PJ, Asner GP, François C and Ustin SL (2009). PROSPECT+ SAIL models: A review of use for vegetation characterization. Remote Sensing of Environment, 113:S56–S66. [7] Féret, J.-B. & de Boissieu, F. An R package for the simulation of canopy reflectance using the model PROSAIL (PROSPECT+SAIL). https://gitlab.com/jbferet/prosail/. [8] Miraglio, T., Adeline, K., Huesca, M., Ustin, S. and Briottet, X. (2022). 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. [9] K. Adeline. APR CNES TOSCA SENTHYMED, REMOTE sensing and in situ observations to study TREEs in natural and manmade landscapes (2023). https://remotetree.sedoo.fr/senthymed/. [10] X. Briottet, K. Adeline, T. Bajjouk, V. Carrère, M. Chami, Y. Constans, Y. Derimian, A. Dupiau, M. Dumont, S. Doz, S. Fabre, P.Y. Foucher, H. Herbin, S. Jacquemoud, M. Lang, A. Le Bris, P. Litvinov, S. Loyer, R. Marion, A. Minghelli, T. Miraglio, D. Sheeren, B. Szymanski, F. Romand, C. Desjardins, D. Rodat, B. Cheul (2024). End-to-end simulations to optimize imaging spectroscopy mission requirements for seven scientific applications. ISPRS Open Journal of Photogrammetry and Remote Sensing, 12, 100060.