131
the hills to be useful. However, in Norfolk, March is often well into the growing
season and a February ‘leaf-off’ image was found to be more useable.
Within the Norfolk study, classification followed a multi-stage stage process: a
draft map was created, which was checked by members of NBIS using randomly
assigned ground survey points (Medcalf et al. 2011). The feedback from NBIS was
used to refine the classification rulebase, especially for features that had not been
well identified in the initial classification. This cycle of refining the maps and field
checking was carried out over several iterations. The final outputs of the mapping
are reported in Medcalf et al. (2013) and show that, following field survey by the
NBIS, the accuracy was found to dependent on the input imagery available. For the
eastern study area, the overall accuracy was 89% but this was lower for the western
study area (78%). The greater classification accuracy in the eastern study area arose
because of the greater temporal spread of imagery available, their higher pixel resolution and greater spectral range.
The errors in the classification are not randomly distributed when the rule based
mapping is used but were spatially concentrated in:
• in areas obscured or shaded by clouds;
• at the boundaries of images; or
• areas with a less than ideal time series of imagery.
At the regional scale in Norfolk, a range of high priority habitats (BAP and
Annex I) were identified using OBIA and rule-based classification (Fig. 6). Splitting
the landscape into its component parts also allowed the classification of habitats,
including the floristically and structurally complex grazing marshes and Breckland
heathlands. Based on the Crick Framework, the priority habitats that were not fully
identified were generally of Type 4a. However, for these habitats, the segmentation
and classification approach was useful for generating ‘areas of search’ as the broader
or ‘parent’ habitats are identifiable and can be delineated. At the landscape scale,
wet heathland types (required for mapping Annex I habitats) could often, but not
always, be mapped with the use of contextual data. The landscape scale work, using
SPOT and IRS imagery, allowed the broad saltmarsh communities (water, sediment,
vegetated saltmarsh) to be separated. However, to distinguish particular components
of vegetation that are relevant to specific priority or Annex 1 habitats within these
more broadly defined classes, high resolution imagery such as GeoEye (1.65 m) and
a high quality Digital Terrain Model (1 m or better) or LiDAR were required.
Sub Regional Mapping: Considerations and Outputs
At a sub-regional scale, finer resolution satellite data are required to map features
such as dykes, small pockets of scrub and wet grasslands. In addition to the satellites
that provided spectral information, LiDAR data was used to give a structural component to the landscape. This allowed the separation of features such as reed beds
Integrated Monitoring for Biodiversity Using Remote Sensing: From Local…
the hills to be useful. However, in Norfolk, March is often well into the growing
season and a February ‘leaf-off’ image was found to be more useable.
Within the Norfolk study, classification followed a multi-stage stage process: a
draft map was created, which was checked by members of NBIS using randomly
assigned ground survey points (Medcalf et al. 2011). The feedback from NBIS was
used to refine the classification rulebase, especially for features that had not been
well identified in the initial classification. This cycle of refining the maps and field
checking was carried out over several iterations. The final outputs of the mapping
are reported in Medcalf et al. (2013) and show that, following field survey by the
NBIS, the accuracy was found to dependent on the input imagery available. For the
eastern study area, the overall accuracy was 89% but this was lower for the western
study area (78%). The greater classification accuracy in the eastern study area arose
because of the greater temporal spread of imagery available, their higher pixel resolution and greater spectral range.
The errors in the classification are not randomly distributed when the rule based
mapping is used but were spatially concentrated in:
• in areas obscured or shaded by clouds;
• at the boundaries of images; or
• areas with a less than ideal time series of imagery.
At the regional scale in Norfolk, a range of high priority habitats (BAP and
Annex I) were identified using OBIA and rule-based classification (Fig. 6). Splitting
the landscape into its component parts also allowed the classification of habitats,
including the floristically and structurally complex grazing marshes and Breckland
heathlands. Based on the Crick Framework, the priority habitats that were not fully
identified were generally of Type 4a. However, for these habitats, the segmentation
and classification approach was useful for generating ‘areas of search’ as the broader
or ‘parent’ habitats are identifiable and can be delineated. At the landscape scale,
wet heathland types (required for mapping Annex I habitats) could often, but not
always, be mapped with the use of contextual data. The landscape scale work, using
SPOT and IRS imagery, allowed the broad saltmarsh communities (water, sediment,
vegetated saltmarsh) to be separated. However, to distinguish particular components
of vegetation that are relevant to specific priority or Annex 1 habitats within these
more broadly defined classes, high resolution imagery such as GeoEye (1.65 m) and
a high quality Digital Terrain Model (1 m or better) or LiDAR were required.
Sub Regional Mapping: Considerations and Outputs
At a sub-regional scale, finer resolution satellite data are required to map features
such as dykes, small pockets of scrub and wet grasslands. In addition to the satellites
that provided spectral information, LiDAR data was used to give a structural component to the landscape. This allowed the separation of features such as reed beds
Integrated Monitoring for Biodiversity Using Remote Sensing: From Local…