Corner Inlet and Nooramunga Habitat Mapping Project Jacquomo Monk 1 , Adam Pope 1 , Daniel Ierodiaconou 1 Kan Otera 2 & Richard Mount 2 1 School of Life and Environmental Sciences, Deakin University (Warrnambool Campus) 2 Blue Wren Group, University of Tasmania November 2011
62
Embed
Corner Inlet and Nooramunga Habitat Mapping Projectdro.deakin.edu.au/eserv/DU:30047216/monk-cornerinlet-2011.pdf · saltmarsh (1,562 ha) habitat classes were most dominant. In addition
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Corner Inlet and Nooramunga Habitat Mapping Project
Jacquomo Monk1, Adam Pope1, Daniel Ierodiaconou1 Kan Otera2 & Richard Mount2
1 School of Life and Environmental Sciences, Deakin University (Warrnambool Campus)
2 Blue Wren Group, University of Tasmania
November 2011
Preferred way to cite: Monk, J. Pope, A. Ierodiaconou, D. Otera, K. Mount, R. (2011). Corner Inlet and Nooramunga Habitat Mapping Project. Deakin University, Warrnambool, Victoria, Australia. 60 pages
Published by the School of Life and Environmental Sciences, Deakin University, Warrnambool, 3280, Australia
The State of Victoria and its suppliers do not warrant the accuracy or completeness of information in this publication and any person using or relying upon such information does so on the basis that the State of Victoria and its suppliers shall bear no responsibility or liability whatsoever for any errors, faults, defects or omissions in the information.
Page | iii
Executive summary
Deakin University and the University of Tasmania were commissioned by Parks Victoria
(PV) to create two updated habitat maps for areas within the Corner Inlet and
Nooramunga Marine and Coastal Park and Ramsar area. The team obtained a ground-
truth data set using in situ video and still photographs. This dataset was used to develop
and assess predictive models of benthic marine habitat distributions incorporating data
from both ALOS (Advanced Land Observation Satellite) imagery atmospherically
corrected by CSIRO and LiDAR (Light Detection and Ranging) bathymetry. This report
describes the results of the mapping effort as well as the methodology used to produce
these habitat maps.
Overall accuracies of habitat classifications were good, returning overall accuracies >73
% and kappa values > 0.62 for both study localities. Habitats predicted with highest
accuracies included Zosteraceae in Nooramunga (91 %), reef in Corner Inlet (80 %), and
bare sediment (no-visible macrobiota/no-visible seagrass classes; both > 76 %). The
majority of classification errors were due to the misclassification of areas of sparse
seagrass as bare sediment. For the Corner Inlet study locality the no-visible macrobiota
(10,698 ha), Posidonia (4,608 ha) and Zosteraceae (4,229 ha) habitat classes covered the
most area. In Nooramunga no-visible seagrass (5,538 ha), Zosteraceae (4,060 ha) and wet
saltmarsh (1,562 ha) habitat classes were most dominant.
In addition to the commissioned work preliminary change detection analyses were
undertaken as part of this project. These analyses indicated shifts in habitat extents in
both study localities since the late 1990s/2000. In particular, a post-classification analysis
Page | iv
highlighted that there were considerable increases in seagrass habitat (primarily
Zosteraceae) throughout the littoral zones and river/creek mouths of both study localities.
Further, the numerous channel systems remained stable and were free of seagrass at both
times. A substantial net loss of Posidonia in the Corner Inlet locality is likely but requires
further investigation due to potential misclassifications between habitats in both the 1998
map (Roob et al. 1998) and the current mapping. While the unsupervised Independent
Components Analysis (ICA) change detection technique indicated some changes in
habitat extent and distribution, considerable areas of habitat change observed in the post-
classification approach are questionable, and may reflect misclassifications rather than
real change. A particular example of this is an apparent large decrease in Zosteraceae and
increase in Posidonia being related to the classification of Posidonia beds as Zosteraceae
in the 1998 mapping. Despite this, we believe that changes indicated by both the ICA and
post-classification approaches have a high likelihood of being ‘actual’ change. A pattern
of gains and losses of Zosteraceae in the region north of Stockyard channel is an example
of this. Further analyses and refinements of approaches in change detection analyses such
as would improve confidence in the location and extent of habitat changes over this time
period.
This work has been successful in providing new baseline maps using a repeatable method
meaning that any future changes in intertidal and shallow water marine habitats may be
assessed in a consistent way with quantitative error assessments. In wider use, these maps
should also allow improved conservation planning, advance fisheries and catchment
management, and progress infrastructure planning to limit impacts on the Inlet
environment.
Page | v
Important note on seagrass species and sediment definition Two species of Zosteraceae were observed in Corner Inlet and Nooramunga; Heterozostera nigricaulis and Zostera muelleri. These, however, could not be consistently differentiated by the remote-sensing techniques employed in this study and have been grouped into a single generic category of ‘Zosteraceae’. As a result, all references to ‘Zosteraceae’ in this report include both Heterozostera nigricaulis, Zostera muelleri and any other species of Heterozostera that may have been present. Additionally, soft sediment type (i.e. sand, silt and mud) could not be reliably differentiated by the remote-sensing techniques used in this study and have been grouped into a ‘no-visible macrobiota’ or ‘no-visible seagrass’ (see Table 3 for class descriptions).
Page | vi
Table of contents
Executive summary ............................................................................................................ iii Table of contents ................................................................................................................ vi List of figures .................................................................................................................... vii List of tables ..................................................................................................................... viii 1 Introduction ................................................................................................................. 9 2 Materials and methods .............................................................................................. 11
2.1 Study locality ...................................................................................................... 11 2.2 LiDAR data ........................................................................................................ 13 2.3 Ground-truth data collection .............................................................................. 14 2.4 Satellite image and atmospheric correction ....................................................... 21 2.5 Habitat map classification .................................................................................. 23
2.5.1 Ground-truth classes ................................................................................... 23 2.5.2 Classification process .................................................................................. 24
2.6 Error Assessment ................................................................................................ 26 2.7 Classification appraisal and quality control ....................................................... 26 2.8 Incorporation of previously mapped intertidal habitats ..................................... 27 2.9 Change detection in seagrass habitats ................................................................ 29
Figure 1. Study Locality. Grey polygons indicate the two study regions within Corner Inlet and Nooramunga Marine and Coastal Park. .......................................................................................... 13
Figure 2. Map for Corner Inlet study locality showing ground-truth points and dominant habitat classes. ............................................................................................................................................ 18
Figure 3. Map for Nooramunga study locality showing ground-truth points and dominant habitat classes. ............................................................................................................................................ 19
Figure 4. Still images of the five dominant habitat types identified and mapped in the two study localities. a) no-visible macrobiota habitat. b) Zosteraceae habitat. c) Posidonia habitat. d) Pyura habitat. e) reef habitat (note: no reef or Pyura habitat was observed in Nooramunga). ................. 20
Figure 5. Differences in ALOS imagery as a result of the atmospheric correction. a) uncorrected. b) AtCor corrected for atmosphere and water column. The corrected image (b) represents the actual spectral response from habitats as would be perceived from immediately above, without the introduced noise and biases from overlying air and water columns. ........................................ 22
Figure 6. Spectral signatures for species that contribute to the major habitats within Corner Inlet and Nooramunga. ........................................................................................................................... 22
Figure 7. Bathymetry and corrected ALOS imagery used in classification process. a) and c) Corner Inlet. b) and d) Nooramunga. Fringing areas in a) and b) reflect clipping of the data to mapped saltmarsh and mangrove distributions. ............................................................................. 25
Figure 8. LandSat scenes showing changes in light and dark patches through time for the Stockyard Channel region of the Corner Inlet study locality. Top: LandSat 5 (29/12/1990). Middle: LandSat 7 (01/01/2000). Bottom: LandSat 5 (10/10/2010) .............................................. 31
Figure 9. Habitat classification map for Corner Inlet study location. ............................................ 37
Figure 10. Habitat classification map for Nooramunga study locality. .......................................... 38
Figure 11. Map of the Stockyard Channel region showing the areas of change delineated by the ICA approach between 2000 and 2010. Red denotes loss. Green denotes gain. ............................ 40
Figure 12. Change detection between 1998 and 2009 maps showing the persistence, loss and gain of grouped ‘seagrass’ in the Corner Inlet study locality ................................................................. 42
Figure 13. Change detection between 1998 and 2009 maps showing the persistence, loss and gain of grouped ‘seagrass’ in the Nooramunga study locality. .............................................................. 42
Figure 14. Post-classification change detection results for Zosteraceae between 1998 and 2009 maps in the Corner Inlet study locality. .......................................................................................... 45
Figure 15. Post-classification change detection results for Posidonia between 1998 and 2009 maps in the Corner Inlet study locality. .......................................................................................... 45
Figure 16. Post-classification change detection results for Zosteraceae between 1998 and 2009 maps in the Nooramunga study locality. ........................................................................................ 46
Figure 17. Post-classification change detection results for Posidonia between 1998 and 2009 maps in the Nooramunga study locality ......................................................................................... 47
Page | viii
List of tables
Table 1. Ground-truth sampling plan. Drops were randomly stratified based on Roob et al. (1998) mapped habitat classes, Parks Victoria Posidonia community monitoring program and ALOS image. ............................................................................................................................................. 15
Table 2. Summary of the number of drops completed within each study location. ....................... 15
Table 3. Dominant biological community selection criteria and ground-truth class size. NVB = no-visible biota dominated. NVSG = no-visible seagrass dominated ............................................ 17
Table 4. Ecological Vegetation Classes (EVC) that were mapped by Boon et al. (2011) and how they were regrouped for the current maps. ..................................................................................... 28
Table 5. Classes used in post-classification change detection. * only mapped in Corner Inlet. ..... 33
Table 6. Depth ranges for the dominant habitat types recorded within each study locality. Depth is in metres relative to lowest astronomical tide (LAT). NVB = no-visible macrobiota. NVSG = no-visible seagrass ............................................................................................................................... 34
Table 7. Error matrix for the Corner Inlet study locality showing the predicted accuracy of each habitat class based on the 25% of ground truth data used for independent assessment. Overall accuracy = 73%; Kappa = 0.62. NVB = no-visible macrobiota ..................................................... 36
Table 8. Error matrix for the Nooramunga study locality showing the predicted accuracy of each habitat class based on the 25% of ground truth data used for independent assessment. Overall accuracy = 85%; Kappa = 0.72. NVSG = no-visible seagrass ....................................................... 36
Table 9. Area of habitat classes in Corner Inlet and Nooramunga based on current map. NVB = no-visible macrobiota; NVSG = no-visible seagrass ...................................................................... 36
Table 10. Comparison of areas of grouped ‘seagrass’ change between 1998 (sourced from Roob et al. 1998) mapping and 2011 mapping for Corner Inlet and Nooramunga study localities derived from ALOS and LiDAR Imagery in this study. ............................................................................. 41
Table 11. Comparison of areas of change (in hectares) between 1998 (sourced from Roob et al. 1998) mapping and 2011 mapping for Corner Inlet and Nooramunga study localities derived from ALOS and LiDAR Imagery in this study. ...................................................................................... 43
Page | 9
1 Introduction
Deakin University and the University of Tasmania were commissioned by Parks Victoria
(PV) to map the distribution of intertidal and shallow water marine habitats in two study
localities within Corner Inlet Ramsar Site. This work was initiated by PV to provide
‘new’ baseline maps so that future changes in intertidal and shallow water marine habitats
linked with catchment processes can be assessed. Additionally, discrepancies between
previous mapping efforts with ground observations by PV staff provided some impetus
for creation of new maps of seagrass extents (pers. comm. Jonathon Stevenson, Parks
Victoria).
Numerous studies have assessed the distribution and degree of change of intertidal and
shallow-water habitats (particularly seagrass habitat) in Corner Inlet and Nooramunga
(Poore, 1978; Morgan, 1986; Jenkins et al. 1993; Allen, 1994; Conron and Coutin, 1995;
Roob et al. 1998; Hindall et al., 2009; Ball et al., 2010). Roob et al. (1998) used
historical aerial photographs to assess the rate of change of seagrass habitat over a 29-
year period and found that there were temporal fluctuations in the level of seagrass cover
for Corner Inlet and Nooramunga; particularly in the northern and north-western regions.
The studies listed above have, however, relied primarily on visual interpretation (and
subsequent manual digitisation) of aerial photography or satellite imagery and limited
ground-truth data to delineate the extents of intertidal and shallow-water habitat within
Corner Inlet and Nooramunga.
Page | 10
Whilst satellite images and aerial photography have been widely used in the mapping and
inventory of coastal resources (Populus and Lantieri, 1990; Zainal et al., 1993; Roob et
al., 1998; Dekker et al., 2003; Mount, 2007; Friedlander et al., 2007; Monk et al., 2008),
recent advances in and access to remotely-sensed bathymetric datasets, such as those
captured using Light Detection and Ranging (LiDAR), provide valuable complementary
information to spectral data for model development. In addition, the resolution and
spectral range of satellites has increased through time providing spectral information that
is important in distinguishing differences in intertidal and shallow-water habitats
(particularly when attempting to separate and map different seagrass species; Chust et al.,
2008). The combination of these datasets, when coupled with advances in geographic
information systems (GISs) and computational power, make it possible to extract novel
spectral and bathymetric information about the seafloor that is important in quantifying
the distribution of shallow water habitats (e.g. Chust et al., 2010).
In contrast to manual digitisation, automated classification techniques (such as the
classification tree approach used in Ierodiaconou et al., 2011 and this study) facilitate the
consistent and repeatable analysis of the large datasets resulting from contemporary
marine remote-sensing technologies and are important in developing reproducible
classification results at scales relevant to management purposes. Automated classification
techniques have also been successfully used to characterise intertidal and shallow-water
substrata and biological habitats using combinations of spectral data (derived from
satellite imagery) and bathymetric information (derived from LiDAR; Chust et al., 2008;
Chust et al., 2010). Spectral information from remote sensing technologies can also be
used as the basis for repeatable change analysis methodologies based on pre-classification
Page | 11
data (Mount et al. 2010). By providing assessments of habitat stability, change detection
provides useful indications of the status and nature of habitat changes through time.
This habitat mapping program combined the recently acquired high-resolution elevation
LiDAR data (collected as a part of the Department of Sustainability and Environment’s
(DSEs) ‘Future Coasts Program’) and ALOS imagery with geo-located ground data in a
GIS-based environment using automated classification techniques. Mapping efforts were
focused on the open-water habitats as the saltmarsh and mangrove habitats have recently
been mapped at sufficient spatial detail under a state-wide initiative (Boon et al. 2011). In
addition to the habitat mapping and beyond the initial scope of the project we also
quantified changes in seagrass habitat using pre and post classification change detection
techniques from 1990 to present to determine the nature of change in seagrass habitat.
2 Materials and methods
2.1 Study locality
The Corner Inlet and Nooramunga bay and barrier island complex covers an area of
approximately 60000 ha and is located approximately 260 km south east of Melbourne in
south eastern Australia, Victoria (Roob et al. 1998; Figure 1).Corner Inlet and
Nooramunga are mostly shallow (> ~ -3 m Lowest Astronomical Tide Datum; LAT) and
have large expanses of intertidal mud and sand flats that are exposed at low tide. Cutting
through these intertidal and mud flats is an extensive network of incised channels (up to
~23 m deep) that drain and fill the inlet complex through its five permanent entrances to
Bass Strait (Roob et al. 1998). Two areas of interest were identified in the brief for this
Page | 12
project, one associated with the western creek system in Corner Inlet and the other with
the estuaries of the Tarra and Albert Rivers in Nooramunga (Figure 1).
The Corner Inlet study locality was situated in the north-western region of the Inlet
(Figure 1) and covered an area of around 21,000 ha. This region is largely devoid of large
islands (with exception of Doughboy Island), being composed almost entirely of mud and
sand flats that support extensive seagrass meadows comprised of Zosteraceae species and
Posidonia australis (Roob et al. 1998). Intermixed within these seagrass meadows are
small patches of ascidians (Pyura sp.), sparse beds of Halophila australis and isolated
sponges. The deeper channels are predominantly comprised of bare sand; although there
are isolated beds of sponges (Roob et al. 1998; O’hara et al. 2002). The littoral zones of
the Corner Inlet study locality are dominated by extensive mangrove forests and
saltmarsh habitats (Boon et al. 2011).
To the east of Corner Inlet is the Nooramunga complex of islands and channels (Figure
1). The Nooramunga study locality covered an area of around 13,000 ha, which is
dominated by many low, sandy islands of various sizes. Some of these islands are ‘barrier
islands’ that form a physical barrier between Bass Strait and the lagoon system of
Nooramunga, and run between Wilson’s Promontory and the Ninety Mile Beach in a
north-easterly direction. Between the islands the mud and sand banks support extensive
seagrass meadows comprised of Zosteraceae species and Posidonia australis (Roob et al.
1998). The Nooramunga study locality supports more extensive mangrove forests than
Corner Inlet. Similar to Corner Inlet, Nooramunga also supports widespread salt marsh
habitats (Boon et al. 2011).
Page | 13
Figure 1. Study Locality. Grey polygons indicate the two study regions within Corner Inlet and Nooramunga Marine and Coastal Park.
2.2 LiDAR data
Bathymetric LiDAR data from the DSE Future Coasts Program were used for the habitat
classification. These data were collected in March-April 2009 using a Hawkeye II ALB
system coupled with a Fugro LADS Mk II inertial motion sensing system and a dual
frequency kinematic geographic positioning system (GPS). LiDAR penetration into the
water column was typically 2-3 times the Secchi depth (Wang and Philpot, 2007). In the
study localities this meant that LiDAR bathymetry was available for almost all of the
region with deep channels being the most common areas without this data (see section
2.5.2).
Page | 14
Flight lines for the mapping survey were spaced at approximately 220 m with a swath
width of 240 m; leaving an overlap of 10 m. Vertical and horizontal accuracy for the
survey were ±0.50 m and ±3.17 m, respectively. Final bathymetry raster grids (elevation
relative to LAT) were gridded at a 10 m pixel resolution using the ArcGIS 9.3 (using bi-
cubic re-sampling) to match the resolution of the ALOS image. In addition, any land
(based on VicMap 1:25000 coastline), saltmarsh and mangrove areas (based on the Boon
et al. 2011 mangrove and saltmarsh GIS layers) were clipped from the bathymetry and
spectral datasets to exclude these regions from the classification analysis.
2.3 Ground-truth data collection
A drop video and still camera system were deployed to collect habitat data with position
provided by differential GPS (Trimble Geo XM) to geo-locate ground-truth records.
Yanakie and Yarram GPS base stations (GPSnet) were used for differential corrections.
For each study locality 30 ground-truth drops were randomly positioned within each of
the dominant habitat types mapped by Roob et al. (1998) (Table 1). Additional Posidonia
localities were selected based on Parks Victoria’s community monitoring program (Table
1). To ensure adequate spatial coverage, additional ground-truth drops were randomly
generated across dark and light patches in the ALOS image (Table 1). Over 8 days
(17/02/11 to 28/02/11) 791 drops were made according to the sampling plan (Table 1).
Fewer drops were achieved than planned (Table 1; Figure 2; Figure 3). This was
predominantly a result of accessibility issues due to tidal movement (for both study
localities), as well as the presence of moored oil rig platform in the eastern section of the
Corner Inlet study locality (Figure 2; Figure 3). Where the bottom could not be clearly
Page | 15
identified from the video drop or still camera, direct in situ observations were made by
snorkelling.
Table 1. Ground-truth sampling plan. Drops were randomly stratified based on Roob et al. (1998) mapped habitat classes, Parks Victoria Posidonia community monitoring program and ALOS image.
Figure 2. Map for Corner Inlet study locality showing ground-truth points and dominant habitat classes.
Page | 19
Figure 3. Map for Nooramunga study locality showing ground-truth points and dominant habitat classes.
Page | 20
Figure 4. Still images of the five dominant habitat types identified and mapped in the two study localities. a) no-visible macrobiota habitat. b) Zosteraceae habitat. c) Posidonia habitat. d) Pyura habitat. e) reef habitat (note: no reef or Pyura habitat was observed in Nooramunga).
a) b)
c) d)
e)
Page | 21
2.4 Satellite image and atmospheric correction
This specifications of this project originally included WorldView-II satellite imagery,
and to this end a tasking order to acquire imagery of the study area was in place from
20/02/2011 to 15/05/2011. Due to the unusually frequent and extensive amount of
cloud cover during this period no suitable imagery was able to be captured. In lieu of
this dedicated imagery an archival scene from ALOS captured on November 8th 2009
was used for the study. This image was selected as it was relatively close in time to
the ground-truthing period, showed very little cloud cover, had a low view angle (0.0
°) and was captured at low tide.
To reduce the effects of atmosphere and water column properties CSIROs AtCor was
applied by CSRIOs Land and Water, Environmental Earth Observation Program
(Figure 5). This process corrects for particulates in the atmospheric (e.g. water
vapour, smoke) and the water column (e.g. suspended organic matter, light
attenuation in the water column). For a detailed description of this process see Brando
et al. (2009). To allow atmospheric correction of the ALOS image, a spectral library
was collected using a spectroradiometer on loan from Geosciences Australia (ASD
FieldSpec® HandHeld spectroradiometer). In total, 331 geo-located (using a
differentially-corrected Trimble Geo XM) spectral signatures from the major habitats
within the two study regions were used for atmospheric correction (Figure 6).
Analysis of the spectral signatures also provided an opportunity to determine the
spectral separability of habitat classes defined and provides a resource for future
assessment of habitat extents in this region.
Page | 22
Figure 5. Differences in ALOS imagery as a result of the atmospheric correction. a) uncorrected. b) AtCor corrected for atmosphere and water column. The corrected image (b) represents the actual spectral response from habitats as would be perceived from immediately above, without the introduced noise and biases from overlying air and water columns.
Figure 6. Spectral signatures for species that contribute to the major habitats within Corner Inlet and Nooramunga.
a) b)
Wavelength (nm)
Page | 23
2.5 Habitat map classification
2.5.1 Ground-truth classes
The no-visible macrobiota habitat class (NVB) was largely devoid of any visible
epibenthic macro fauna or flora (i.e. < less than 10 % cover; Table 3; Figure 4). In
Nooramunga the Pyura class was too small to effectively separate it from the no-
visible macrobiota class (i.e. large amount of misclassification between the two
classes). Accordingly, for this locality no-visible macrobiota and Pyura classes were
clumped and termed ‘no-visible seagrass’ dominated habitat. This habitat class was
largely devoid of any visible seagrass (i.e. < less than 10 % cover; Table 3; Figure 4),
but in some instances other species (e.g. Pyura sp.) were observed within images at a
range of densities. The Zosteraceae, Posidonia and Pyura habitat classes were
characterised by > 25 % cover by their respective species. In some instances these
classes also had < 25% cover of the other species observed within the images (see
Table 3; Figure 4). The reef dominated habitat class was defined by >25% cover of
reef (irrespective of the biological habitat; Table 3; Figure 4). Hormosira banksii,
Halophila australis and sponges were also observed within some images, but usually
not at a high enough abundance to form a dominant habitat type. Consequently, these
species were grouped into one of the five dominant habitat types.
Classified ground-truth data were converted to point data in a GIS (ArcGIS 9.3) and
translated from a from geographic World Geodetic System projection (WGS 1984) to
the Cartesian Geocentric Datum of Australia 1994, Map Grid of Australia zone 55
projection (GDA 94 MGA zone 55) using the bi-cubic re-sampling in ArcGIS 9.3.
The data were then converted to an ESRI ASCII grid format (10 m cell size)
compatible with ENVI 4.8 (ITTVIS Inc.) remote sensing software. A stratified
random sampling method (for dominant biota classes) was used to divide the ground-
Page | 24
truth datasets for training and evaluation datasets. Seventy five percent of the ground-
truth data were used as training regions in the thematic classification. The remaining
25% of ground–truth data that were not used in the classification were set aside for a
subsequent accuracy assessment that would not be biased by data used to train the
classification.
2.5.2 Classification process
Since the LiDAR datasets had some regions with missing data (caused by high
turbidity levels) a ‘holes filling’ process was used to replace the ‘missing data’
regions with values based on surround cells (Figure 7a,b). The process was
undertaken in ENVI 4.8 and provided complete coverage of both study localities. The
masked LiDAR and ALOS imagery were stacked within ENVI 4.8. These stacked
datasets were then combined with the ground-truth training datasets to enable
prediction of habitats. A Quick Unbiased Efficient Statistical Tree approach (QUEST;
Loh & Shih 1997) was applied to each study locality separately to predict the spatial
distribution of habitat classes based on the ground-truth observations. The QUEST
approach was executed using Rule Gen v1.02 extension in ENVI 4.8. For model runs,
class-prior probabilities (i.e. the estimated probability that an observation belongs to
that particular class) were based on training sample size for each category, scaled
from 0 to 1. Classification-tree models are vulnerable to overfitting, where the model
reflects the structure of the training data set too closely. Even though a model appears
to be accurate on training data, if overfitted, it may be much less accurate when
applied to a wider data set. To account for overfitting a ten-fold cross-validation
pruning with one standard error rule and minimum node size of 5 samples were used
on the classification trees.
Page | 25
Figure 7. Bathymetry and corrected ALOS imagery used in classification process. a) and c) Corner Inlet. b) and d) Nooramunga. Fringing areas in a) and b) reflect clipping of the data to mapped saltmarsh and mangrove distributions.
Page | 26
2.6 Error Assessment
The remaining 25% of ground-truth observations (independent of those used to
develop the classification) were compared with derived thematic classifications to
produce two forms of error assessment. First, the results were used to construct error
matrices showing map accuracy measures for individual classes within biota and
substrate categories. Second, overall accuracy was derived by calculating the
percentage of correctly classified pixels. The Kappa coefficient of agreement (Khat)
was used to derive a measure of accuracy between the classified map and the
independent ground-truth data. By including errors of omission and commission in the
calculations, kappa analysis takes into account errors expected by chance (Foody
2002, Jensen 2005). The Khat coefficient of agreement gives a more accurate overall
representation of the accuracy of the thematic mapping exercise and allows better
comparison with error matrices derived from other survey areas. We used Fleiss's
(1981) guidelines to characterize kappa values: > 0.75 as excellent, 0.40-0.75 as fair
to good, and <0.40 as poor.
2.7 Classification appraisal and quality control
Following the automated classification process, results for the two study localities
were appraised relative to existing knowledge of the study areas and known
distributions of habitat by a local expert (J. Stevenson) as well as the project team.
This assessment looked for presence of obvious LiDAR data artefacts (related to the
holes-filling process of original LiDAR dataset) and any misclassification of satellite
imagery. Any demonstrably incorrectly classified (i.e. areas with ground observation
contrary to classified habitat types) were then contextually edited to the correct class
using an analysis mask within the ArcGIS Spatial Analyst raster calculator.
Common misclassification issues included:
Page | 27
1. Some deeper areas within channel systems were being classified as Pyura
habitat instead of no-visible macrobiota. Based on ground observations, Pyura
in the study area predominantly occurred on the edge of sand banks with
isolated patches in the channel systems.
2. Areas where bathymetry data were missing and were substituted with dummy
data resulted in artefacts (striping) that were often misclassified as Posidonia.
2.8 Incorporation of previously mapped intertidal habitats
The final intertidal and subtidal habitat classifications were combined with the
saltmarsh and mangrove datasets surveyed by Boon et al. (2011). Using the classes
described in that work there were a total of 33 different saltmarsh and mangrove
communities mapped in the study areas (Table 4).
At the spatial resolution of the current mapping effort (i.e. ~1:19,000) many of these
communities were not visible and arguably not sufficiently distinctive in the context
of the range and resolution of habitat types being considered. Consequently, these
communities were regrouped into broader communities of no-visible macrobiota, wet
saltmarsh, dry saltmarsh and mangroves (Table 4). For applications where knowledge
of particular saltmarsh community classes at a finer spatial scale is required the
original GIS layers from Boon et al. (2011) should be used. Regrouped classes from
this earlier mapping were combined with intertidal and subtidal habitat classes from
this project to produce a map for each study locality.
Page | 28
Table 4. Ecological Vegetation Classes (EVC) that were mapped by Boon et al. (2011) and how they were regrouped for the current maps.
Original EVC Grouped EVC Locality
Bare Sand Bank No-visible macrobiota
Nooramunga
Bare Sediment No-visible macrobiota
Both
Coastal Dry Saltmarsh Dry Saltmarsh Both
Coastal Dry Saltmarsh/Estuarine Wetland Dry Saltmarsh Both
Two different methods were used to assess change in seagrass habitat through time:
(1) an unsupervised image differencing approach (i.e. independent component
analysis; ICA) using archived LandSat imagery at a decadal scale, and (2) a post-
classification approach assessing differences between Roob et al. 1998 map and the
current classifications. These two approaches were selected as they are
complementary and have slightly different strengths and weaknesses. The two
approaches are summarised below and described in detail in sections 2.9.1 and 2.9.2.
ICA approach:
Pros
• May extract maximum change: based on all variation in the spectral dataset
• Includes reference for change, so change is anchored at starting value, unlike
change vector analysis and image differencing
Cons
• May be extremely difficult to interpret classes
Post-classification approach:
Pros
• Avoids need for strict radiometric calibration
• Favours classification scheme of user
• Designates type of change occurring
Cons
• Error is propagated from two parent maps
• Changes within classes are not detected
Page | 30
2.9.1 Independent component analysis
Independent Component Analysis (ICA) is a statistical and computational technique
for linear transformation (Hyvärinen 1999; Hyvärinen and Oja 2000). ICA attempts to
detect a linear representation of non-Gaussian data for extracting original signals,
which are statistically independent or as independent as possible from each other
(Hyvärinen and Oja 2000; Robila et al. 2000). Where such a linear representation
delineates a specific underlying pattern in the data, then ICA is useful for feature
extraction from the data (Hyvärinen and Oja 2000).
For satellite remote sensing, ICA can be used for image feature extraction based on
the spectral characteristics of multi or hyper-spectral images (Robila et al. 2000).
Often in coastal regions, like this study area, image classification is difficult due to
spectral confusion with adjacent land cover classes (Ozesmi and Bauer 2002). ICA
has the potential to distinguish classes with spectral similarity as statistically
independent. For more detail on the ICA process see Otera (2009).
Three results were produced using the ICA approach for the Stockyard Channel
region of Corner Inlet for three time steps over 20 years:
(1) Dec 1990 and Jan 2000,
(2) Jan 2000 and Oct 2010, and
(3) Dec 1990 and Oct 2010.
The ICA routine in ENVI version 4.7 was applied to the single stacked LandSat
dataset for each time step. Each resultant dataset from the routine delineates an
independent component of the original data, such as changed and unchanged areas of
vegetation between time periods, or individual feature of time 1 or time 2. The quality
of extracted features highly depends on the spectral response of input dataset. If the
Page | 31
dataset contains a large amount of noise or systematic error, these features could be
extracted instead of the targeted features such as seagrass beds. As a consequence of
sensor noise in the 1990 image (see 1990 image in Figure 8), only the post 2000
results (time step 2) are presented in this report.
Figure 8. LandSat scenes showing changes in light and dark patches through time for the Stockyard Channel region of the Corner Inlet study locality. Top: LandSat 5 (29/12/1990). Middle: LandSat 7 (01/01/2000). Bottom: LandSat 5 (10/10/2010)
Page | 32
2.9.2 Post-classification approach
Post-classification change detection makes direct comparison between the results of
mapping efforts with similar habitat classifications and can be used regardless of the
methods used to derive these classifications. An underlying assumption of post-
classification is that both maps being compared are ‘accurate’ at the time they were
made. In this case, the 1998 map was known to have had some misclassifications of
Posidonia habitat as Zosteraceae (J. Stevenson pers. comm.) but thefull extent of
these misclassifications is unknown – this exercise is likely best estimate of the scale
of misclassification. Consequently, Zosteraceae and Posidonia habitat were grouped
into a general ‘seagrass’ category for comparison purposes. The other classes present
in either map (e.g. Pyura in the current map) were grouped and termed ‘other’ (Table
5).
While the grouped comparison is useful as is mitigates the misclassification errors in
classes and is comparable with the ICA approach, a second post-classification
comparison was also undertaken. This second comparison compared Zosteraceae and
Posidonia change separately to quantify change, where possible, in these two
ecologically distinct habitat classes.
For both comparisons the 1998 map was clipped to the same extent as the current
map. Habitat classes in the 1998 map were grouped to reflect similar classes as
mapped in this report (Table 5).Using the tabulate areas tool in ArcGIS 9.3, mapped
class areas were cross tabulated to enabled the identification of persistence, gain and
loss of habitat classes between the 1998 map (Roob et al. 1998) and the present map
for the two study localities. Net change is the difference in area of a habitat class
between time 1 and time 2. Gain refers to the increase in area of a habitat class, while
loss refers to a decrease in area of a habitat class between time 1 and time 2.
Page | 33
Table 5. Classes used in post-classification change detection. * only mapped in Corner Inlet.
1998 Map Current Map Original class Reclass for
change detection
Original class Reclass for change detection
Bare Bare Bare (NVB/NVSG) Bare Dense Posidonia Posidonia Zosteraceae Zosteraceae
Medium Posidonia Posidonia Posidonia Posidonia Sparse Posidonia Posidonia Reef* Other Sparse Posidonia & Halophila mix
Dense Halophila Other Intertidal Vegetation Other Land Other Sparse Halophila Other
Page | 34
3 Results
3.1 Depth range summaries for intertidal and subtidal habitat classes
The depth distributions for each habitat class are shown in Table 6. As reflected by
the standard deviation of depth, Posidonia exhibited minimal variation in depth for
both study localities, and was constrained to regions > -2.5 m and < -0.7 m (Table 6).
Similarly, Zosteraceae was confined to shallow regions (i.e. > -6.5 m), but exhibited
slightly greater variation in depth extending into the deep intertidal (Table 6). Pyura
and no-visible macrobiota/seagrass had the greatest variation in depth ranges down to
~ - 18 m. The reef class was confined to intertidal and very shallow subtidal regions
(i.e. ~ 0.2 m).
Table 6. Depth ranges for the dominant habitat types recorded within each study locality. Depth is in metres relative to lowest astronomical tide (LAT). NVB = no-visible macrobiota. NVSG = no-visible seagrass
Corner Inlet Mean depth (Stand. Dev.)
Min. depth Max. depth
NVB -2.86 (4.09) 1.62 -17.98
Zosteraceae -1.02 (1.05) 0.26 -6.53
Posidonia -1.32 (0.33) -0.68 -2.28
Pyura -3.96 (4.50) -1.03 -17.28
Reef 0.20 (0.36) 0.69 -0.50
Nooramunga NVSG -1.73 (2.36) 1.45 -10.01
Zosteraceae -1.21 (0.96) 0.02 -4.95
Posidonia -1.61 (0.47) -1.18 -2.55
3.2 Intertidal and subtidal habitat classification
The classification procedure provided good predictions of the dominant habitat types
observed within the two study localities, returning overall accuracies of 73 % (kappa
0.62) for Corner Inlet and 85 % (kappa 0.72) for Nooramunga (Table 7; Table 8).
Page | 35
3.2.1 Corner Inlet Locality
For the Corner Inlet locality, the no-visible macrobiota category (i.e. bare sand or
mud) covered 50 % of the site (Table 9), with this habitat being mainly confined to
the deeper channel systems (Figure 9). Posidonia-dominated habitat was mapped over
22 % of the study locality. In the north this habitat was primarily confined to edges of
sand banks but it dominated the shallow sand banks to the south-east of the area
(Figure 9). Zosteraceae-dominated habitat covered 20 % of the study region and was
primarily confined to the sand and mud flats close to the shoreline (Figure 9). The
remaining intertidal and subtidal habitats (e.g. saltmarsh, mangroves, Pyura and reef
dominated habitats) were estimated to cover a combined total of 8.9 % of the study
were primarily confined to the littoral zone of the Inlet (Figure 9). Pyura was
primarily confined to the edges of sand banks and the in channel systems (Figure 9).
3.2.2 Nooramunga Locality
Similar to Corner Inlet, a large proportion (41 %) of the Nooramunga study locality
had zero or very low seagrass cover (no-visible seagrass category) (Table 9; Figure
10). Zosteraceae dominated habitat covered 30 % of the study locality and was mainly
confined to the northern, more sheltered, regions (Figure 10). The remaining intertidal
and subtidal habitats (e.g. woodland, saltmarsh, mangroves and Posidonia-dominated
habitats) were predicted to cover a combined total of 10 % of the study locality (Table
9; Figure 10). The largest areas of Posidonia were located on flats near the outer
channels (Figure 10). Wet saltmarsh and mangroves were more prominent at the
Nooramunga than the Corner Inlet site, covering approximately 12 % and 10 % of the
study locality respectively (Table 9; Figure 10).
Page | 36
Table 7. Error matrix for the Corner Inlet study locality showing the predicted accuracy of each habitat class based on the 25% of ground truth data used for independent assessment. Overall accuracy = 73%; Kappa = 0.62. NVB = no-visible macrobiota
Table 8. Error matrix for the Nooramunga study locality showing the predicted accuracy of each habitat class based on the 25% of ground truth data used for independent assessment. Overall accuracy = 85%; Kappa = 0.72. NVSG = no-visible seagrass
Table 9. Area of habitat classes in Corner Inlet and Nooramunga based on current map. NVB = no-visible macrobiota; NVSG = no-visible seagrass
Corner Inlet Nooramunga Habitat Area (ha) Percent
age Area (ha) Percentage
Woodland 0 0 22 0.2Dry Saltmarsh 29 0.1 482 3.6
Wet Saltmarsh 424 2.0 1562 11.6
Saline Wetland 14 0.1 107 0.8
Mangrove 843 4.0 1329 9.9
Reef 51 0.2 0 0.0Zosteraceae 4229 19.7 4060 30.3
Posidonia 4608 21.5 317 2.4
Pyura 553 2.6 0 0.0NVB 10698 49.9 0 0.0
NVSG 0 0 5538 41.3
Total 21436 100 13132 100
Page | 37
Figure 9. Habitat classification map for Corner Inlet study location.
Page | 38
Figure 10. Habitat classification map for Nooramunga study locality.
Page | 39
3.3 Change detection
3.3.1 Independent component analysis
The results of the ICA approach indicated some spectral changes in areas of known
seagrass habitat in the Corner Inlet study area between 2000 and 2010 (Figure 11). In
some areas spectral patterns associated with areas of seagrass had changed to those
associated with bare sediment (depicted in red; Figure 11), while the reverse was true
for other areas (depicted in green; Figure 11). The largest area of gain was on
sandbanks north of Stockyard channel with a band of loss areas further inshore.
Small areas of gain were also recorded in the northwest of the area, near the mouths
of Old Hat, Stockyard and Bennison creeks. Elsewhere a mixed pattern of losses and
gains was recorded. As the ICA was an unsupervised approach (i.e. dark pixels were
inferred to be seagrass of any species, while light pixels were inferred to be bare),
results were compared with an analogous post-classification change detection
approach using all seagrass categories combined from the 1998 and current habitat
maps.
Page | 40
Figure 11. Map of the Stockyard Channel region showing the areas of change delineated by the ICA approach between 2000 and 2010. Red denotes loss. Green denotes gain.
3.3.2 Post-classification analyses
Post-classification comparisons of combined seagrass categories between the 1998
map and the current mapping effort are summarised in Table 10; Figure 12 and Figure
13. In the Corner Inlet study locality 59 % of combined seagrass habitat mapped in
1998 was also mapped as seagrass in this study (Table 10; Figure 12). A substantial
proportion of the total area of seagrass was recorded in different locations to those in
1998 with 41% of the area mapped as seagrass in 1998 apparently lost since then but
offset by newly mapped beds covering a greater area inshore (equivalent to 56% of
the total 1998 area). These apparent gains were predominantly observed around the
littoral zone of the Inlet with a mixed pattern of gains and losses across the tops of
Page | 41
sandbanks further into the inlet. Patterns of loss and gain in the Corner Inlet locality
shared some similarities with the ICS results, notably large areas of gain on the
sandbanks around the Stockyard channel and the smaller patches of loss immediately
to the northwest of the larger gain areas. A notable difference between the pre- and
post-classification analyses was the absence of substantial gains in littoral regions
with the ICA comparison.
Sixty seven percent of the mapped area of combined seagrass classes in Nooramunga
remained unchanged between the 1998 and current mapping, with a considerable
overall increase to an area 388% larger than the seagrass habitat mapped in 1998
(Table 10; Figure 12). These apparent gains are located primarily in littoral areas
along the coastlines of the mainland and islands. Apparent losses are around the main
channel on the west of Sunday Island with scattered areas of apparent loss elsewhere.
Table 10. Comparison of areas of grouped ‘seagrass’ change between 1998 (sourced from Roob et al. 1998) mapping and 2011 mapping for Corner Inlet and Nooramunga study localities derived from ALOS and LiDAR Imagery in this study.
Figure 12. Change detection between 1998 and 2009 maps showing the persistence, loss and gain of grouped ‘seagrass’ in the Corner Inlet study locality
Figure 13. Change detection between 1998 and 2009 maps showing the persistence, loss and gain of grouped ‘seagrass’ in the Nooramunga study locality.
Page | 43
A second post-classification change detection was undertaken to determine the
amount of change in Zosteraceae and Posidonia classes between 1998 and the current
map. Table 11 shows changes in habitat classes for the two study localities between
1998 and the current map. There were substantial large differences in area and losses
and gains were apparent for all classes. The classes with the largest proportion of
persistent habitat between the two mapping events were Bare (63%) and Posidonia
(58%) classes in the Corner Inlet site and Bare (57%) and Zosteraceae (45%) classes
in the Nooramunga study locality. Other classes had low (~15% or less) persistence
between the two mapping dates.
Table 11. Comparison of areas of change (in hectares) between 1998 (sourced from Roob et al. 1998) mapping and 2011 mapping for Corner Inlet and Nooramunga study localities derived from ALOS and LiDAR Imagery in this study.
Study Locality
Habitat Persistence (ha)
Loss (ha) Gain (ha)
Corner Inlet Bare (NVB) 7773 4564 2885 Zosteraceae 915 5183 3315 Posidonia 949 687 3659 Other 4 22 597 Total 9641 10455 10456 Nooramunga Bare (NVSG) 4805 3617 376 Zosteraceae 512 608 3547 Posidonia 1 7 316 Other 0 6 0 Total 5318 4238 4239
Page | 44
An overall decrease in bare area was apparent in both localities, accompanied by
increases in the area of seagrass. In the Corner Inlet locality there was a net loss of
30% in the area covered by Zosteraceae and an increase of 81% in the area covered by
Posidonia. These changes are due to a combination of:
• gains in Zosteraceae in the nearshore region (Figure 14);
• losses of Zosteraceae in the area immediately seaward of the above gains
(Figure 14);
• substantial areas identified in this survey as Posidonia which were classified
as Zosteraceae by Roob et al. (1998) and so were recorded as gains and losses
in each category respectively (Figure 14, Figure 15); and
• patchy losses of Posidonia adjacent to existing beds (Figure 15).
There was also a substantial increase of 575ha in the area covered by the ‘Other’ class
(Table 11). This was most likely related to the differences in classes used between the
1998 and current studies, and the reclassification used for change detection (Table 5).
Differences driving this result relate to the presence of ‘reef’ and ‘Pyura’ classes in
this study but not the 1998 results and so apparent losses of the ‘bare’ class being
replaced by gains in the ‘other’ class.
Page | 45
Figure 14. Post-classification change detection results for Zosteraceae between 1998 and 2009 maps in the Corner Inlet study locality.
Figure 15. Post-classification change detection results for Posidonia between 1998 and 2009 maps in the Corner Inlet study locality.
Page | 46
In Nooramunga there was a large increase in the area mapped as Zosteraceae between
the 1998 and 2009 mapping. This was largely related to gains in the nearshore area
(Figure 16). Losses of Zosteraceae in this study locality area appeared to be primarily
related to an area which changed from the Zosteraceae class to the Posidonia class
between the two mapping exercises (Figure 16, Figure 17). Apparent gains in
Posidonia for the Nooramunga study site were largely due to the change in
classification of this area.
Figure 16. Post-classification change detection results for Zosteraceae between 1998 and 2009 maps in the Nooramunga study locality.
Page | 47
Figure 17. Post-classification change detection results for Posidonia between 1998 and 2009 maps in the Nooramunga study locality
4 Discussion
The aims of this study were to map the distribution of intertidal and subtidal marine
habitats for two study localities within Corner Inlet and Nooramunga Ramsar area in a
digital form at a nominal scale of ~1:25,000 or larger. Given that the saltmarsh and
mangrove habitats have recently been mapped by Boon et al. (2011), the current
mapping effort was directed at mapping intertidal and subtidal bare sediment (no-
visible macrobiota/seagrass), seagrass and invertebrate habitats. These were later
combined with saltmarsh and mangrove datasets to create an overall map for the two
study localities.
Page | 48
Based on the two derived maps it was apparent that the most widely distributed
habitats were no-visible macrobiota (50 %; Corner Inlet) and the equivalent no-visible