Predicting ecosystem services in agricultural woodlands from airborne hyperspectral images - INRAE - Institut national de recherche pour l’agriculture, l’alimentation et l’environnement Accéder directement au contenu
Communication Dans Un Congrès Année : 2017

Predicting ecosystem services in agricultural woodlands from airborne hyperspectral images

Résumé

Ecosystem services has become a major sociological, ecological and economic issue worldwide. Woodlands of agricultural landscapes provide valuable services although there are often underestimated and poorly used. Identifying and mapping these services at large scale is an important step to locate the available resources and to plan future management. However, accurate mapping of these services remains challenging. Remotely-sensed data enable to cover large area and to describe in a novel way agricultural landscapes, offering a possibility to predict ecosystem services at wide scale. This seems particularly relevant, as ecosystem services of isolated, small agricultural woodlands may strongly depends on landscape-level processes. We assessed the contribution of vegetation indices derived from high spatial resolution hyperspectral images in predicting ecosystem services provided by agricultural woodlands, as compared to predictions based on land cover. 28 woodland patches were sampled for supporting, regulating and producing ecosystem services. Airborne hyperspectral images were acquired for the study area, at a 2m resolution. Usually, hyperspectral-derived vegetation indices are reduced (many pixels to plot or landscape-level indices) using descriptive statistics (mean, standard deviation, minimum and maximum values). However, this approach may lose a lot of information, especially at landscapes scale, where several objects with various spectral signatures are present. We propose a hyperspectral landscape description based on the full distribution of vegetation indices across landscapes. We used gaussian mixture models (gmm) of pixel distributions within each landscapes, and introduced a L2 distance based on these approximated distributions. Finally, we used a k-nearest neighbors approach to predict ecosystem services (similarity-based prediction). The quality of predictions were compared across three landscape representations: land-cover, descriptive statistics of hyperspectral data, and gmm-based description of hyperspectral data. Preliminary results suggest that gmm are much better proxies than descriptive statistics for estimating the distributions of vegetation indices across landscapes and seems to be better predictors of ecosystem services.
Fichier non déposé

Dates et versions

hal-02733689 , version 1 (02-06-2020)

Identifiants

  • HAL Id : hal-02733689 , version 1
  • PRODINRA : 409644

Citer

Rémi Duflot, Aude Vialatte, David Sheeren, Mathieu Fauvel. Predicting ecosystem services in agricultural woodlands from airborne hyperspectral images. IUFRO 8.01.02 Landscape Ecology Conference 2017, Sep 2017, Halle, Germany. 140 p. ⟨hal-02733689⟩
14 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More