19
and unknown flaws and anomalies,…), commonly applied pre-processing algorithms may further inflate differences, even between two images of the same sensor
(through resampling and reductions in spatial, spectral and/or radiometric resolution) (Jones and Vaughan 2010). For vegetation monitoring however, constraints in
applications most often arise from difficulties in obtaining suitable high-resolution
imagery. A limited number of data providers, cloudy weather, and yearly variations
in vegetation phenology, make it very difficult to obtain truly comparable pairs of
image data from two different timestamps.
Real-World Examples
Next, we present some examples of transfer cases from our own experience, illustrating the type of problems that occurred, and the solutions that were (sometimes)
found.
The method we tested for transferability was originally developed in the framework of a Belgian research project called ‘HABISTAT’. Its aim was to map Natura
2000 habitat types and their conservation status in Natura 2000 sites, with a focus
on Atlantic heathland in Flanders. The method is described in Haest et al. (2010)
and Haest et al. (2017). In short, it consists of a supervised classification (Linear
Discriminant Analysis) of vegetation using classes geared towards remote identification, followed by a rule-based reclassification into occurrences (‘patches’) of
Natura 2000 habitat.
In the frame of a follow-up project ‘MS.MONINA’ (EU 7th Framework
Programme), three transfer cases were tested, with the method being applied to: (a)
the original study area, but at later dates, (b) a Continental heathland ecosystem in
Germany, and (c) a coastal dune ecosystem in Flanders. In the course of the tests,
several method adjustments had to be made:
• In the original context of the HABISTAT project, the study area of Atlantic
heathland was intensively visited, with many hundreds of field reference plots
recorded in either carefully selected homogeneous vegetation patches or in random locations. These served as training and validation data respectively. In a
more operational remote sensing context, such large datasets are rarely available,
and therefore, the method had to be extended to accommodate different types of
training data. In particular, the method was adapted to work with extracted training data from existing (field-driven) vegetation maps, rather than the point-based
field plots in the original version. We considered two options, using either single
or multiple points per polygon of the field map.
• For the new study sites in the transfer cases, a new remote-sensing oriented
classification scheme had to be developed, based on vegetation descriptions or
maps of the new study sites. It was not always straightforward to ‘translate’
existing field data into the new classification scheme, either because the existing
classification was too detailed or not detailed enough, or because the existing
field information was biased towards a few specific vegetation types only.
Towards a Mature Age of Remote Sensing for Natura 2000 Habitat Conservation: Poor…
and unknown flaws and anomalies,…), commonly applied pre-processing algorithms may further inflate differences, even between two images of the same sensor
(through resampling and reductions in spatial, spectral and/or radiometric resolution) (Jones and Vaughan 2010). For vegetation monitoring however, constraints in
applications most often arise from difficulties in obtaining suitable high-resolution
imagery. A limited number of data providers, cloudy weather, and yearly variations
in vegetation phenology, make it very difficult to obtain truly comparable pairs of
image data from two different timestamps.
Real-World Examples
Next, we present some examples of transfer cases from our own experience, illustrating the type of problems that occurred, and the solutions that were (sometimes)
found.
The method we tested for transferability was originally developed in the framework of a Belgian research project called ‘HABISTAT’. Its aim was to map Natura
2000 habitat types and their conservation status in Natura 2000 sites, with a focus
on Atlantic heathland in Flanders. The method is described in Haest et al. (2010)
and Haest et al. (2017). In short, it consists of a supervised classification (Linear
Discriminant Analysis) of vegetation using classes geared towards remote identification, followed by a rule-based reclassification into occurrences (‘patches’) of
Natura 2000 habitat.
In the frame of a follow-up project ‘MS.MONINA’ (EU 7th Framework
Programme), three transfer cases were tested, with the method being applied to: (a)
the original study area, but at later dates, (b) a Continental heathland ecosystem in
Germany, and (c) a coastal dune ecosystem in Flanders. In the course of the tests,
several method adjustments had to be made:
• In the original context of the HABISTAT project, the study area of Atlantic
heathland was intensively visited, with many hundreds of field reference plots
recorded in either carefully selected homogeneous vegetation patches or in random locations. These served as training and validation data respectively. In a
more operational remote sensing context, such large datasets are rarely available,
and therefore, the method had to be extended to accommodate different types of
training data. In particular, the method was adapted to work with extracted training data from existing (field-driven) vegetation maps, rather than the point-based
field plots in the original version. We considered two options, using either single
or multiple points per polygon of the field map.
• For the new study sites in the transfer cases, a new remote-sensing oriented
classification scheme had to be developed, based on vegetation descriptions or
maps of the new study sites. It was not always straightforward to ‘translate’
existing field data into the new classification scheme, either because the existing
classification was too detailed or not detailed enough, or because the existing
field information was biased towards a few specific vegetation types only.
Towards a Mature Age of Remote Sensing for Natura 2000 Habitat Conservation: Poor…