Informations et ressources scientifiques
sur le développement des zones arides et semi-arides

Accueil du site → Doctorat → Italie → Mediterranean Forest Species Mapping Using Hyperspectral Imagery

Università degli studi di Trieste (2013)

Mediterranean Forest Species Mapping Using Hyperspectral Imagery

Galidaki, Georgia

Titre : Mediterranean Forest Species Mapping Using Hyperspectral Imagery

Auteur : Galidaki, Georgia

Université de soutenance : Università degli studi di Trieste

Grade : Tesi di dottorato 2013

Advances in hyperspectral technology provides scientists the opportunity to investigate problems that were difficult if not impossible to approach using multispectral data ; among those, species composition which is a very important and dynamic forest parameter, linked with many environmental qualities that we want to map and monitor. This study addresses the problem of Mediterranean forest species mapping using satellite EO-1 Hyperion imagery (30m, 196 bands). Two pixel based techniques were evaluated, namely Spectral Angle Mapper (SAM) and Support Vector Machines (SVM), as well as an object oriented approach (GEOBIA). These techniques were applied in two study areas with different species composition and pattern complexity, namely Thasos and Taksiarchis. Extensive field work provided reference data for the accuracy assessment of the produced maps. Image preprocessing included several steps of data corrections and the Minimum Noise Fraction transformation, as means for data dimensionality reduction. In the case of Thasos, where two conifer species are present, SAM technique resulted in an overall accuracy (OA) of 3.9%, SVM technique yielded OA of 89.0% and GEOBIA achieved an OA of 85.3%. In the case of Taksiarchis, where more species are present – both conifers and broadleaved- the respective OA was 80.0%, 82.6% and 74.1%. All three methodologies implemented to investigate the value of hyperspectral imagery in Mediterranean forest species mapping, achieved very accurate results ; in some cases equivalent to forest inventory maps. SAM was the straightest forward to implement, only depending on the training samples. Implementation SVM involved the specification of several parameters as well as the use of custom software and was more successful in the challenging landscape of Taksiarchis. GEOBIA adapted to scale through segmentation and extended the exercise of classification, allowing for knowledge based refinement. Lower accuracies could be attributed to the assessment method, as research on alternative assessment methods better adapted to the nature of object space is ongoing. Two typical Mediterranean forests were studied. In Thasos, two conifer species of the same genus, namely Pinus brutia and Pinus nigra, dominate a big part of the island. Both of them were accurately mapped by all methodologies. In Taksiarchis primarily stands of Quercus frainetto mix with stands of Fagus sylvatica and the aforementioned pines. The two pines were again mapped with high accuracy. However, there was a notable confusion between the two broadleaved species, indicating the need for further research, possibly taking advantage of species phenology. The outcome of the proposed methodologies could confidently meet the current needs for vegetation geographical data in regional to national scale, and demonstrate the value of hyperspectral imagery in Mediterranean forest species mapping.

Mots clés : hyperspectral ; forest species mapping ; SAM ; SVM ; GEOBIA ; Hyperion ; Remote sensing


Version intégrale (4,47 Mb)

Page publiée le 30 juin 2013, mise à jour le 18 mars 2019