Estimating leaf mass per area and equivalent water thickness based on leaf optical properties: Potential and limitations of physical modeling and machine learning

Feret Jean Baptiste, Le Maire Guerric, Jay Sylvain, Berveiller Daniel, Bendoula Ryad, Hmimina Gabriel, Cheraiet A., Oliveira J.C., Ponzoni Flávio Jorge, Solanki T., De Boissieu Florian, Chave Jérôme, Nouvellon Yann, Porcar-Castell A., Proisy Christophe, Soudani Kamel, Gastellu-Etchegorry J.P., Lefèvre-Fonollosa Marie-José. 2019. Estimating leaf mass per area and equivalent water thickness based on leaf optical properties: Potential and limitations of physical modeling and machine learning. Remote Sensing of Environment, 231:110959, 14 p.

https://doi.org/10.1016/j.rse.2018.11.002

Article de revue ; Article de recherche ; Article de revue à facteur d'impact

Version publiée - Anglais
Accès réservé aux personnels Cirad
Utilisation soumise à autorisation de l'auteur ou du Cirad.
2019Feret_RSE_LMA_EWT.pdf
Télécharger (2MB) | Demander une copie

Quartile : Outlier, Sujet : ENVIRONMENTAL SCIENCES / Quartile : Q1, Sujet : REMOTE SENSING / Quartile : Outlier, Sujet : IMAGING SCIENCE & PHOTOGRAPHIC TECHNOLOGY

Liste HCERES des revues (en SHS) : oui

Thème(s) HCERES des revues (en SHS) : Géographie-Aménagement-Urbanisme-Architecture

Résumé : Leaf mass per area (LMA) and leaf equivalent water thickness (EWT) are key leaf functional traits providing information for many applications including ecosystem functioning modeling and fire risk management. In this paper, we investigate two common conclusions generally made for LMA and EWT estimation based on leaf optical properties in the near-infrared (NIR) and shortwave infrared (SWIR) domains: (1) physically-based approaches estimate EWT accurately and LMA poorly, while (2) statistically-based and machine learning (ML) methods provide accurate estimates of both LMA and EWT. Using six experimental datasets including broadleaf species samples of >150 species collected over tropical, temperate and boreal ecosystems, we compared the performances of a physically-based method (PROSPECT model inversion) and a ML algorithm (support vector machine regression, SVM) to infer EWT and LMA based on leaf reflectance and transmittance. We assessed several merit functions to invert PROSPECT based on iterative optimization and investigated the spectral domain to be used for optimal estimation of LMA and EWT. We also tested several strategies to select the training samples used by the SVM, in order to investigate the generalization ability of the derived regression models. We evidenced that using spectral information from 1700 to 2400 nm leads to strong improvement in the estimation of EWT and LMA when performing a PROSPECT inversion, decreasing the LMA and EWT estimation errors by 55% and 33%, respectively. The comparison of various sampling strategies for the training set used with SVM suggests that regression models show limited generalization ability, particularly when the regression model is applied on data fully independent from the training set. Finally, our results demonstrate that, when using an appropriate spectral domain, the PROSPECT inversion outperforms SVM trained with experimental data for the estimation of EWT and LMA. Thus we recommend that estimation of LMA and EWT based on leaf optical properties should be physically-based using inversion of reflectance and transmittance measurements on the 1700 to 2400 nm spectral range.

Mots-clés Agrovoc : indice de surface foliaire, biomasse, végétation, spectroscopie infrarouge, modélisation environnementale, informatique, propriété optique, risque d'incendie, prévention des incendies, réduction des risques d'incendie, apprentissage machine