Research project S0/02/074 (Research action S0)
Context and objectives
The research aims to map saltmarsh vegetation for inventory, management and monitoring purposes in respect to nature reserves and other (semi)-natural areas. The vegetation maps will show the extent, type and change of vegetation types and their biophysical/biochemical characteristics. Maps can therefore be used as indicators of the ecological properties of the study area (Island of Schiermonnikoog, The Netherlands). At the same time the maps are means of satisfying the demand by managers and policy makers for environmental geo-information.
The main scientific objective is to develop object-based approaches for hyperspectral image analysis that facilitate the management of these threatened ecosystems. From the research project, we expect several important outcomes: (i) increased scientific knowledge and expertise in the field of imaging spectroscopy, mandatory for the utilization of the APEX instrument and succeeding satellite payloads, (ii) new and improved algorithms for reliable hyperspectral data processing, (iii) thematic maps of the study area for use in ecosystem management, and (iv) the establishment of an international network including scientific and public partners from Belgium and The Netherlands.
Methodology
The object-based approach for image segmentation/classification and radiative transfer model (RTM) inversion can be summarized by the following general workflow: (1) segmentation of the hyperspectral image data into image-objects (2) extraction of object-signatures using previously delineated object boundaries (3) supervised classification (ML) of different vegetation types based on previously extracted object-signatures. Due to the large number of spectral variables contained in the object-signatures, a MNF-transformation will be applied prior to classification. Training samples will be taken from field plot data with known vegetation types (4) object-based inversion of SAILH+PROSPECT radiative transfer model (RTM) and retrieval of biophysical (leaf area index, percent cover) and biochemical variables (leaf chlorophyll/nitrogen content) (5) accuracy assessment of the classification results (7) comparison of the object-based approach with conventional “baseline” approaches relying only on a subset of the full spectrum. For the “baseline” classification a conventional ML-classifier will be used. (8) upscaling of the results to MERIS and CHRIS image scales, and extrapolation over the Belgium and Netherlands coastlines
Results
The following results were obtained from the (still ongoing) project:
1) A comprehensive reference dataset based on fieldwork
2) a workflow for pre-processing and correcting hyperspectral imagery
3) A pixel based classification strategy
4) Coupling of genetic algorithms to the spectral angle mapper for the selection of appropriate input bands
5) A segmentation approach targeted at hyperspectral remotely sensed data
6) Establishment of a relation between segment orientation and humidity / salinity gradients
All of the above was implemented in either c++ or IDL/ENVI
Furthermore, we expect to be able to produce a feature selection approach in combination with artificial neural network object based classification in the near future. The latter is subject of ongoing research.
