25
Transfer Test Case 3: To a Different Ecosystem –
from Heathlands to Coastal Dunes
Study Site
The Flemish nature reserve De Westhoek is one of the oldest nature reserves in
Belgium and the largest remaining area of coastal dunes in the country, bordering
France at its westernmost tip. It hosts a variety of Natura 2000 habitats, such as
Marram (Ammophila arenaria) dunes (2120), grey dunes (2130), Sea-buckthorn
(Hippophae rhamnoides) shrub (2160), Creeping willow (Salix repens) shrub
(2170), wooded dunes (2180), and humid dune slacks (2190). Although this array of
habitats is completely different from Kalmthoutse Heide, the landscape is structurally very similar, consisting mainly of dry sandy dunes, either bare or covered with
grasses, mosses or low shrubs, interspersed with humid depressions and occasional
snippets of woodland. Moreover, there are also parallels in threats and disturbances
to both sites: dune fixation, invasive exotic species, groundwater extractions and
recreational use are affecting the functioning of the ecosystem. Therefore, we considered it appropriate to test the transferability of our method from heathland ecosystems to coastal dune ecosystems.
Method Application
Hyperspectral APEX images of the area were acquired on 14 June 2011. The first
task was to draw up a remote sensing oriented classification scheme for this site.
This was especially crucial since it was the first time the method was to be applied
outside heathlands. For the Westhoek, a detailed (almost to the level of dominant
species) and recent (2010) vegetation map was available, which proved very valuable for this task. The resulting classification scheme is shown in Table 3.
Training data extraction proved less straightforward. Since the source data were
provided as polygons, selecting random pixels within these polygons for training
could not exclude the possibility of including impure pixels. Furthermore, the
training dataset was highly imbalanced, with a high amount of Hawthorn (Crataegus
monogyna). The first trial classification runs indeed showed this imbalance to cause
problems, but after several adaptations (see introduction to section “Real-world
examples”), the accuracy levels could be brought to an acceptable level (see Table 4;
validation based on a separate dataset extracted from the same source map of 2010).
However, the next step, the translation into a habitat map by the rule-based reclassification algorithm (Haest et al. 2017), resulted in clear errors: the ‘No habitat type’
class dominated overall although, in reality, the area is almost entirely covered by
Natura 2000 habitats. It seems that the higher species diversity in coastal dunes,
compared to heathlands, proved difficult to accommodate in simple rules. Substantial
additional fine-tuning of the rule-base would be needed to make this step work satisfactorily, which we did not pursue, as it would have taken more time and raised the
risk of over-fitting to one site. The classification result at Level 3 and an excerpt of
the erroneous habitat map is shown in Fig. 5.
Towards a Mature Age of Remote Sensing for Natura 2000 Habitat Conservation: Poor…
Transfer Test Case 3: To a Different Ecosystem –
from Heathlands to Coastal Dunes
Study Site
The Flemish nature reserve De Westhoek is one of the oldest nature reserves in
Belgium and the largest remaining area of coastal dunes in the country, bordering
France at its westernmost tip. It hosts a variety of Natura 2000 habitats, such as
Marram (Ammophila arenaria) dunes (2120), grey dunes (2130), Sea-buckthorn
(Hippophae rhamnoides) shrub (2160), Creeping willow (Salix repens) shrub
(2170), wooded dunes (2180), and humid dune slacks (2190). Although this array of
habitats is completely different from Kalmthoutse Heide, the landscape is structurally very similar, consisting mainly of dry sandy dunes, either bare or covered with
grasses, mosses or low shrubs, interspersed with humid depressions and occasional
snippets of woodland. Moreover, there are also parallels in threats and disturbances
to both sites: dune fixation, invasive exotic species, groundwater extractions and
recreational use are affecting the functioning of the ecosystem. Therefore, we considered it appropriate to test the transferability of our method from heathland ecosystems to coastal dune ecosystems.
Method Application
Hyperspectral APEX images of the area were acquired on 14 June 2011. The first
task was to draw up a remote sensing oriented classification scheme for this site.
This was especially crucial since it was the first time the method was to be applied
outside heathlands. For the Westhoek, a detailed (almost to the level of dominant
species) and recent (2010) vegetation map was available, which proved very valuable for this task. The resulting classification scheme is shown in Table 3.
Training data extraction proved less straightforward. Since the source data were
provided as polygons, selecting random pixels within these polygons for training
could not exclude the possibility of including impure pixels. Furthermore, the
training dataset was highly imbalanced, with a high amount of Hawthorn (Crataegus
monogyna). The first trial classification runs indeed showed this imbalance to cause
problems, but after several adaptations (see introduction to section “Real-world
examples”), the accuracy levels could be brought to an acceptable level (see Table 4;
validation based on a separate dataset extracted from the same source map of 2010).
However, the next step, the translation into a habitat map by the rule-based reclassification algorithm (Haest et al. 2017), resulted in clear errors: the ‘No habitat type’
class dominated overall although, in reality, the area is almost entirely covered by
Natura 2000 habitats. It seems that the higher species diversity in coastal dunes,
compared to heathlands, proved difficult to accommodate in simple rules. Substantial
additional fine-tuning of the rule-base would be needed to make this step work satisfactorily, which we did not pursue, as it would have taken more time and raised the
risk of over-fitting to one site. The classification result at Level 3 and an excerpt of
the erroneous habitat map is shown in Fig. 5.
Towards a Mature Age of Remote Sensing for Natura 2000 Habitat Conservation: Poor…