Spectrometry for plant species recognition in Mediterranean forest canopies - Spectral separability and image unmixing studied on field and aerial data

Abstract

Mediterranean-type forests are formed by heterogeneous, dense communities. This condition represents a challenge for plant species discrimination using optical remote sensing. The detection of species based on top-canopy reflectance is hampered by the fact that spectral signatures of different crowns/species mix up in the spectrum sampled by the sensor. The thesis develops a method which integrates field and aerial hyperspectral data together with a ground validation dataset. Object-based image analysis and spectral unmixing techniques are used to estimate the species fraction cover.

The study area is located in Hérault province – France; the hilly landscape of former farmlands presents a continuous forest cover. The research considers the six arboreal species which are regularly found in the forest canopies, throughout the whole study area. The fieldwork component of the project was carried out in September-October 2011. Crown and leaf reflectance were measured in-situ with a spectrometer. The species composition and other characteristics of the canopy were recorded at sampling plots.

As an upgraded method for species mapping, the thesis investigates the viability of introducing “mixed” classes in the image spectral mixture analysis. The mixed classes are identified on the basis of the ground observations and are generated from the “pure” ones retrieved by the spectrometer. A statistical analysis on the field spectra precedes the image analysis. Statistic tests serve to study the spectral discriminability of the species and their mixed combinations; the tests also isolate a subset of bands used in the image analysis. The crown spectra produce the endmembers for the image spectral mixture analysis. Pure and mixed endmembers are both used. Linear Spectral Unmixing is applied to a previously segmented HyMap image. The accuracy assessment is performed on the basis of the sampling plots. The species fraction cover observed on the ground is compared with that estimated on the reflectance.

The input of mixed endmembers in the image mixture analysis results in a smaller estimation error on average. The thesis concludes that using mixed endmembers is a valuable option to improve the species mapping in a heterogeneous forest.