Remote Sensing of Environment 86 (2003) 480 – 490
www.elsevier.com/locate/rse
Mapping intertidal estuarine sediment grain size distributions
through airborne remote sensing
M.P. Rainey a, A.N. Tyler a,*, D.J. Gilvear a, R.G. Bryant b, P. McDonald c
a
Department of Environmental Science, University of Stirling, Stirling, Scotland FK9 4LA, UK
b
Department of Geography, University of Sheffield, Sheffield S10 2TN, UK
c
Westlakes Scientific Consulting Ltd, Westlakes Science and Technology Park, Moor Row, Cumbria CA24 3LN, UK
Received 22 October 2001; received in revised form 1 May 2003; accepted 3 May 2003
Abstract
The intertidal environments of estuaries represent critical exchange environments of sediment and sediment bound contaminants.
Ecological and sedimentological related investigations of these environments require monitoring methods that provide rapid spatially
representative data on sediment grain size distribution. Remote sensing has the potential to provide synoptic information of intertidal
environments. Previous in situ and laboratory-based reflectance investigations have demonstrated that for effective quantification of sediment
grain size distributions, remote sensing platforms must include measurements within the short-wave infrared (SWIR). In addition, the timing of
image acquisition, in relation to tidal cycles and sediment moisture content, is critical in optimising the spectral differences between the coarser
sand and finer ‘mud’ fraction of sediments. Daedalus 1268 Airborne Thematic Mapper (ATM) has been identified as an appropriate platform
and sensor for providing accurate synoptic maps of estuarine sediment distributions. This paper presents the results from the application of
ATM 1.75 m resolution data to the mapping of surface sediment grain-size distributions across intertidal areas of Ribble Estuary, Lancashire,
UK. ATM imagery was acquired after the intertidal area was exposed to strong summer drying conditions. Pre-processing and linear unmixing
of the imagery collected of the intertidal zone following a period of drying allowed accurate sub-pixel determinations (1.75 m resolution) of
sediment clay (r2 = 0.79) but less accurate for sand (r2 = 0.60). The results also demonstrate deterioration in the image calibration with
increasing sediment moisture content and microphytobenthos cover. However, recombining the subpixel end member abundances through
multivariate regression analysis improved the image calibration significantly for both sediment clay and sand content (r2>0.8) for imagery
collected in both dryer and wetter conditions. These results demonstrate that ATM data, or similar, can be used to gain quantitative information
on intertidal sediment distributions and such data has application to a wide variety of estuarine research.
D 2003 Elsevier Inc. All rights reserved.
Keywords: Estuaries; Sediments; Airborne remote sensing; Linear mixture modelling
1. Introduction
Estuaries are critical interfaces between the marine and
terrestrial environment. The deposition of the fine ‘mud’
fraction occurs as a result of the interaction between currents,
tides and salinity. Microphytobenthos is also an important
ecological mechanism contributing to the distribution of fine
sediment through processes of sediment sequestering and
stabilisation (Black & Paterson, 1997; Paterson, 1997).
Estuaries therefore represent a restricted exchange environ-
* Corresponding author. Tel.: +44-1786-467838; fax: +44-1786467843.
E-mail address: [email protected] (A.N. Tyler).
0034-4257/03/$ - see front matter D 2003 Elsevier Inc. All rights reserved.
doi:10.1016/S0034-4257(03)00126-3
ment that can act as a sink or a source of sediment and
sediment associated contaminants discharged to the aquatic
environment. Knowledge of the distribution of fine sediment
can provide an important insight into the understanding of the
fate of sediment bound pollutants discharged to the marine
environment. For example, anthropogenic radionuclides discharged under license into the Irish Sea from British Nuclear
Fuels (BNFL) Sellafield plant since 1952 can be detected in
several UK coastal and estuarine environments (Cook,
MacKenzie, McDonald, & Jones, 1997; Mackenzie, Cook,
McDonald, & Jones, 1998; Tyler, 1999). Consequently, there
is a growing interest by regulatory and industrial bodies into
monitoring particle bound contaminants.
It is well recognised that estuarine environments tend to
exhibit considerable spatial and temporal heterogeneity.
M.P. Rainey et al. / Remote Sensing of Environment 86 (2003) 480–490
Characterisation of this heterogeneity through isolated point
samples is commonly time-consuming, expensive and often
unrepresentative (Rainey, 1999; Tyler, Sanderson, Scott, &
Allyson, 1996). It would therefore be advantageous to a
range of environmental disciplines to develop a remote sensing approach, which could provide rapid quantitative synoptic data of the intertidal sediment grain size distributions.
The ability to map intertidal surfaces through airborne
remote sensing has been demonstrated using Compact Airborne Spectrographic Imager (CASI) data (Thomson, Eastwood, Yates, Fuller, Wadsworth, & Cox, 1998; Thomson,
Fuller, Sparks, Yates, & Eastwood, 1998) and simulated
Daedalus 1268 Airborne Thematic Mapper (ATM) data
(Bryant et al., 1996). However, the application of remote
sensing to the quantitative mapping of sediment grain size
distributions has had limited success. Previous studies have
shown that a ‘soft’ or fuzzy classification technique is likely
to give a more accurate result than ‘hard’ classification
techniques (Donoghue, Reid Thomas, & Zong, 1994; Yates,
Jones, McGrorty, & Goss-Custard, 1993). Of the available
‘soft’ techniques, linear mixture modelling was considered
the most suitable, primarily because: (1) it does not require
extensive training data; (2) it has been shown to be relatively
successful in unmixing satellite data of similar environments
(Mertes, Smith, & Adams, 1993; Yates et al., 1993); (3) the
necessary software is easily accessible. However, the success
of this previous work has been somewhat limited by factors
controlled by the temporal differences between image acquisition and reference data collection. Bryant et al. (1996)
demonstrated the importance the SWIR to distinguish the
influence of moisture on intertidal sediment reflectance.
Rainey, Tyler, Bryant, Gilvear, and McDonald (2000)
showed, through in situ and laboratory-based reflectance
experiments, that image acquisition following a period of
drying would be critical in maximising the spectral distinction between the coarser sand and finer ‘mud’ dominated
sediment fractions. The spectral influence introduced by
481
grain size itself, as demonstrated in areas with otherwise
relatively uniform surface characteristics (Painter, Roberts,
Green, & Dozier, 1998), in this context is likely to be of
minor importance other than to facilitate drainage and drying
of sediment thereby promoting the spectral contrast between
sand and mud. Principal components analysis of simulated
ATM data derived from the reflectance spectra indicated that
linear mixture modelling of ATM imagery acquired of the
intertidal zone, following a period of drying, is a suitable
method for providing relative sub pixel abundance estimates
of sediment grain size distributions (Rainey et al., 2000). The
resulting relative abundance images when coupled with
reference data for calibration should enable quantitative
mapping of sediment grain size distributions.
Following the findings of Rainey et al. (2000), this paper
presents the results from an investigation that examines the
capability of airborne remote sensing, in this case ATM
imagery, to quantitatively map the surface characteristics of
intertidal sediments.
2. Study area
The Ribble Estuary in Lancashire, UK, was selected for
the development of the application of airborne remote sensing
to map intertidal sediment grain-size distributions, because
access was good to the intertidal environment and the range
of sediment grain sizes were present. The Ribble Estuary
joins the Irish Sea at Lytham St. Annes (Fig. 1). The estuary is
macrotidal, experiencing diurnal tides of up to 10 m and the
normal tidal limit reaches above Preston Docks (Brown,
1997). In the late 19th and early 20th century, extensive
engineering work was carried out to improve navigation in
the estuary, which resulted in a relatively straight, narrow
channel and relatively rapid sediment infilling (Van der Wal,
Pye, & Neal, 2002). Training walls were constructed facing
the main channel position and consequently these allowed
Fig. 1. Outline of the 1997 ATM image acquisition strategy and the location of the associate sampling points.
482
M.P. Rainey et al. / Remote Sensing of Environment 86 (2003) 480–490
mudflats and saltmarsh regions to form. Since the early
1980s, dredging of the estuary ceased, resulting in further
extensive siltation of the channel (HMIP, 1994).
The fluvio-glacial deposits on the floor of the Irish Sea are
the dominant source of sediment to the Ribble Estuary,
although the input of fine sediments are thought to be mainly
alluvial in origin. Whole-rock mineralogical analysis of a
range of intertidal sediment samples determined that all the
samples were dominantly quartz with associate plagioclase,
orthoclase, calcite, dolomite, chlorite/kaolinite and mica
(Brown, 1997; Bryant et al., 1996). Clay mineralogical
analysis determined that illite was the dominant clay mineral,
with smectite, kaolinite and chlorite present as minor components (Brown, 1997; Bryant et al., 1996). Throughout the
estuary, variability in the mineralogical composition of the
fine sediment was found to be low, due to the sediments
having the same source and subsequent homogenisation
through reworking and mixing (Brown, 1997). Halite is
occasionally present as an efflorescent deposit on the polygonated dry mud (Bryant et al., 1996).
3.1.2. Reference data
A total of 77 sediment samples were collected from three
intertidal locations within the Ribble Estuary during the
acquisition of the imagery (Fig. 1). The locations of the
individual sample sites were defined using in situ orientation
measurements, differential GPS and marker boards that were
visible in the imagery. To investigate the error introduced to
the ground truth results by sampling and analytical error, 10
samples were collected from each of three 2 2 m quadrats
situated between the saltmarsh edge and the main channel at
Warton Bank.
3.2. Image preparation
3.2.1. Radiometric and geometric correction
Radiometric correction is an essential part of the ATM
image preparation and was performed externally by Natural
Environment Research Council (NERC) of the United Kingdom (Wilson, Mockeridge, & White, 1996). Subsequent,
geometric correction reduced any spatial errors within the
raw ATM imagery and this was achieved using the GCORR
software provided by NERC.
3. Methods
3.1. Reference data collection and ATM image acquisition
3.1.1. Image dataset
A Piper PA31 350 Chieftain aircraft, operated by the UK’s
Natural Environment Research Council (NERC), was the
platform used to collect imagery of the Ribble Estuary on
May 30, 1997 (Fig. 1). The Daedalus 1268 ATM onboard the
aircraft is a passive sensor designed to collect and record
radiation from the earth’s surface. The radiation is separated
into 11 spectral bands, which simulate the satellite-borne
LANDSAT Thematic Mapper and range from the visible
blue to the thermal infrared. Channels 1– 5 are in the visible,
6 –8 are in the near-infrared, 9 and 10 are in the short wave
infrared (SWIR), and lastly, band 11 is a thermal infrared
band. The pixel resolution is 1.75 m and the swath width is 1
km at 700 m altitude. The Wild RC-10 camera, which has a
6-in. lens and a navigation sight, was used to collect vertical
colour stereoscopic photographs in coincidence with the
ATM image collection.
The image dataset contains five east – west orientated
flight lines of ATM imagery of the Ribble Estuary collected
following a neap tide on May 30, 1997, when the intertidal
area was completely exposed (from approximately 09:54 to
15:54). Based on the in situ and laboratory reflectance
measurements it was expected that the imagery collected
after the longest period of exposure would provide the best
results. Consequently, the last or ‘dry’ image was used to
produce a grain-size distribution map for the entire intertidal
area of the Ribble Estuary. A subset of the first or ‘wet’
image was also analysed to examine the influence of interstitial moisture on the ability to create grain-size maps from
ATM imagery.
3.2.2. Atmospheric correction
Atmospheric correction attempts to remove the spectral
effects of the atmosphere, ensuring the ATM recorded
radiance is as close as possible to the actual radiance reflected from the ground surface. The imagery was collected
at 700 m during uniform cloud-free conditions. No pathlength effects were identified and a simple dark pixel
subtraction, or histogram minimum method, was selected
as a suitable and reliable atmospheric correction technique
to remove the effects of scattering or haze from the
imagery.
3.2.3. Image subsetting and masking
To classify accurately the sediments within the intertidal
areas, it was first necessary to minimise the spectral complexity of the image by isolating the intertidal regions from
all other spectral features. Consequently, all non-intertidal
areas such as the surrounding saltmarsh, the main channel
and the man-made structures were identified through the
associated colour aerial photographs and masked out of the
imagery.
3.3. Image analysis and interpretation
The methodology used to create grain-size distribution
maps from ATM images of intertidal sediment has a number
of discrete stages:
(1) Maximisation of the signal-to-noise ratio, as a means of
improving the final classification products.
(2) Examination of each dataset in feature space and,
subsequent, identification of the image spectral endmembers.
M.P. Rainey et al. / Remote Sensing of Environment 86 (2003) 480–490
(3) Classification of the imagery based on the ‘pure’ image
pixels representing each of the spectral end-members.
(4) Calibration of the abundance imagery produced by linear unmixing.
3.3.1. Minimum noise fraction transform
The minimum noise fraction (MNF) transformation is a
noise reduction technique designed to maximise the signalto-noise ratio and hence improve the image quality (Green,
Berman, Switzer, & Graig, 1988). Each transformation of a
10-band ATM image, excluding the thermal band, produces
10 MNF bands of decreasing image information. This allows
bands consisting predominantly of noise to be excluded from
subsequent analysis. The spectral dimensionality of the
dataset, which is fundamental to linear spectral unmixing,
may also be determined.
The first 10 ATM bands of each masked image were
separately placed through a forward MNF transform, producing 10 bands of decreasing image quality with increasing
component number and an accompanying eigenvalue graph.
Following examination of the eigenvalues and eigenimages
of each band, those containing a significant amount of
coherent image information were selected as the basis for
spectral end-members selection and image unmixing. To
validate the noise-removal ability of this technique, the
coherent MNF bands, produced from the transform of the
‘dry’ image, were placed through an inverse MNF transform
and the resultant images were compared to the original
masked imagery.
3.3.2. Linear mixture modelling
Linear spectral unmixing attempts to calculate the
proportion of the various surface components present in
each image pixel based on the spectral characteristics of
the surfaces. The number of spectral end-members used in
linear unmixing must be less than or equal to one greater
than the number of spectral dimensions in the data (Settle
& Drake, 1993).
The number of coherent MNF bands produced from each
image was used as an approximation of the dimensionality of
each dataset and, consequently, to estimate the number of
spectral end-members that could be unmixed. The coherent
MNF bands were plotted against each other in scatterplots to
examine the distribution of the data in feature space and the
nature of the spectral mixing. From the scatterplots the
spectral end-members of each scene were identified, using
a similar approach to Adams, Smith, and Gillespie (1993).
The extreme of each end-member in feature space (ca. 12 –
20 pixels) was subsequently highlighted and the pure representative image pixels were identified. These image pixels
were then selected and a mean spectrum was calculated for
each end-member. The source of each spectral end-member
was identified from a combination of aerial photographs, in
situ knowledge and the mean spectra plots. The mean endmember spectra were used as the basis for ‘unconstrained’
spectral unmixing of the coherent MNF bands. This tech-
483
nique was chosen in preference to ‘constrained’ unmixing
due to the potentially non-linear spectral nature of the
moisture end-member in the intertidal environment (Rainey
et al., 2000).
The end-member pixel selection and unmixing procedure
described above was an iterative process. Initially, pure
pixels were selected to represent the dominant spectral
end-members of each scene and used to unmix the imagery.
The subsequent abundance images were then assessed using
in situ knowledge of the area and the associated root mean
square (RMS) error image. On this basis, it was determined
whether the spectral end-members of the imagery were
adequately represented by the selected end-member pixels.
If this did not prove to be the case, the selection of endmember pixels was refined and the spectral unmixing of the
imagery was repeated. In general, the linear unmixing of an
image was found to be insensitive to variations in the
selection of pixels representing distinct spectral end-members. However, where the spectral end-member was indistinct, minor variations in the selected representative pixels
resulted in substantial variations in the resulting abundance
imagery. As a solution to this problem, various combinations
of the coherent MNF bands were plotted against each other
until the extreme of the spectral end-member was better
defined, making it easier to select representative pixels.
3.3.3. Calibration
As detailed, sediment samples were collected during the
image acquisition to enable calibration of the subsequent
sediment abundance maps. The sample sites were located in
the imagery from white marker boards and sample site
distance measurements. The sand, mud, water and microphytobenthos abundance of each sample site was calculated
from the images produced by ‘unconstrained’ linear unmixing. Regression analysis was undertaken using Mintab vs.
13. Where appropriate, best-subsets analysis was undertaken
to identify the optimum multivariate regression relationship
between the sample sediment characteristics derived from
ground sampling and the abundance estimates derived from
the linear mixture modelling. The final calibration derived
from the multivariate regression analysis was subsequently
used to produce maps of percentage clay and sand abundance
across the intertidal area.
4. Results
4.1. Image preparation
The image preparation process successfully produced a
registered image of the exposed intertidal zones present
within the ‘dry’ image and the ‘wet’ image subset, in the
correct geometric orientation. The dark pixel subtraction
technique resulted in no significant visual changes in the
imagery. No path-length differences in image radiance values were observed.
484
M.P. Rainey et al. / Remote Sensing of Environment 86 (2003) 480–490
4.2. MNF transform
The MNF transform of the masked ‘dry’ image produced
10 bands, of which the first four contained 93% of the total
statistical variance in the 10-band image dataset and are
represented by coherent eigenimages. Similarly, examination of the eigenvalues of the MNF bands produced by the
transform of the ‘wet’ subset reveals that 95% of the image
information is held within the first four bands of the
imagery. For both images, the MNF Band 5 eigenimage
was found to contain relevant information and was combined with the first four bands, and used as the basis of the
end-member selection and unmixing process. The remaining
incoherent bands were discarded. From the associated
eigenvalues, it was determined that the images had approximately four data dimensions, suggesting that five endmembers could be successfully unmixed from each image.
The coherent MNF bands produced from the ‘dry’ image
underwent an inverse MNF transform and the resulting
image was subsequently compared to the original ATM
image. Visual inspection revealed that the original ATM
bands have a higher signal-to-noise ratio, especially ATM
Band 1. This suggests that the MNF transform successfully
separated the majority of the image noise into the incoherent
MNF bands without removing a significant amount of image
information.
4.3. End-member selection
From both the ‘dry’ and ‘wet’ images, four dominant
spectral end-members were identified in the imagery from
2D scatterplots of various combinations of the first four
MNF bands (Fig. 2). These were subsequently found to
represent mud, sand, water and microphytobenthos present
in the imagery. In Scatterplot 1 (Fig. 2A), the data are very
compressed in feature space with only one obvious endmember, sand. Pure sand and water image pixels are more
clearly identifiable in Scatterplot 2 (Fig. 2B). However, the
mud end-member is masked by the microphytobenthos endmember in both Scatterplots 1 and 2, and is only clearly
differentiated in Scatterplot 3 (Fig. 2C) where they have
unique positions in feature space relative to the MNF Band 4.
4.4. Pure end-member image pixels
For both images, the mud spectral end-member was
represented by pure image pixels corresponding to areas of
polygonated mud close to the edge of the saltmarsh. Pure
sand pixels were identified on the ridge of a large sandbank
west of Warton in the outer estuary. Areas within a large
creek and channel water in the outer estuary represented the
water end-members of both images. A small area of
unmasked saltmarsh vegetation at the edge of a large creek
from Warton Bank represented the extreme of the vegetation/
microphytobenthos end-member. The mean spectra of these
representative pure pixels were found to agree in general
Fig. 2. 2D scatterplots of: (A) MNF Band 1 vs. MNF Band 2, (B) MNF Band
2 vs. MNF Band 3, and (C) MNF Band 3 vs. MNF Band 4. The dominant
spectral end-members are highlighted.
M.P. Rainey et al. / Remote Sensing of Environment 86 (2003) 480–490
485
and the actual clay content of the sample sites (r2 = 0.79,
p < 0.01).
%Clay ¼ exp½1:39 þ 1:44Md Fig. 3. The mean spectra of the end-member pixels used to unmix the ‘dry’
image.
with the in situ spectral knowledge of these surface types
(Fig. 3) (Rainey et al., 2000). For example, in the higher
bandwidths, moisture has an overall low reflectance, and
sand has a substantially higher overall reflectance than mud.
The mean spectra of the end-member pixels are found to be
nearly identical for both the ‘dry’ and ‘wet’ images.
4.5. Intertidal mud distribution
4.5.1. ‘Dry’ ATM image
The selected end-member pixels were used to linearly
unmix the intertidal areas of the ‘dry’ image, producing sand,
water, mud and microphytobenthos abundance maps, and an
associated RMS error image. The mud abundance image
appears to be an good representation of the mud distribution
within the intertidal zones of the Ribble Estuary (Fig. 4).
This is supported by the empirical relationship (given in Eq.
(1)) between the image-derived mud abundance estimates
ð1Þ
where Md is the mud abundance estimate from the ‘dry’
image.
Fig. 5, illustrates the relationship between the Mud
Abundance estimate and percent clay content derived from
surface sediment scrapings. This demonstrates that sediment
clay content can be indirectly calculated from the mud
abundance imagery produced by spectral unmixing. This
relationship is based on 77 sample sites from three contrasting locations in the Ribble Estuary and is therefore assumed
to be representative for the whole intertidal area.
The associated RMS error image demonstrates that the
unmixing error is relatively low within the fine-grained
intertidal sediment areas. However, the spectral profiles
reveal that where microphytobenthos cover is high, the
mud abundance appears to be slightly underestimated
(Fig. 6). In order to correct this, a multivariate regression
was performed to incorporate the influence from the microphytobenthos abundance in addition to the mud abundance
estimates. The empirical relationship between the abundance estimates and the observed clay content was improved (r2 = 0.815, p < 0.01), supporting the suggestion
that microphytobenthos attenuates the spectral characteristics of mud (Eq. (2)). The result of the regression model
is compared to the original sediment data in Fig. 7.
%Clay ¼ exp½1:56 þ 1:04Md þ 0:677Bd ð2Þ
where Bd is the microphytobenthos abundance estimate
from the ‘dry’ image.
The exponential models described by Eqs. (1) and (2)
provide the optimum relationships between image abundance and sample-derived results. Whilst linear unmixing
Fig. 4. Subset of the mud abundance image produced from the ‘dry’ image.
486
M.P. Rainey et al. / Remote Sensing of Environment 86 (2003) 480–490
Fig. 5. Regression plot of the relationship between mud abundance and clay
content derived from sediment samples (r2 = 0.79, p < 0.01).
was used, the relationship between mud abundance and clay
content is clearly non-linear. The use of natural logarithms
also normalises the data distribution for the percentage clay
content results satisfying the conditions for regression analysis. Fig. 7 justifies the comparison of the model and sample
derived percent clay content, and also demonstrates likely
magnitude of the sampling error (1 standard deviation)
derived from Table 1. This demonstrates that much of the
scatter about the regression line may be accounted for by
sampling error and that even though ‘unconstrained’ spectral
unmixing was used, the final calibrated results do no exceed
the physical bounds of 0 –100% clay content.
Similarly, surface water is found to mask the true spectral
characteristics of the underlying mud, causing underestimation (Fig. 8). High interstitial moisture contents, however,
appears to only influence the reflectance characteristics of
areas where the mud signal is very weak, i.e. the actual clay
levels are lower than 2%. Finally, the channel walls and the
Fig. 7. Comparison between image predicted clay content (Eq. (2)) and clay
content derived from sediment samples (r2 = 0.815, p < 0.01). Error bars
indicate sampling error, expressed as 1 standard deviation.
stone-topped large ridges, found in a number of locations in
the Ribble Estuary, e.g. Longton area, were misclassified as
areas of high mud abundance, probably due to the similar
spectral nature of these surface types. Mud draping may be a
contributing factor to this miss classification.
4.5.2. ‘Wet’ ATM image
As with the ‘dry’ image, the mud abundance estimates
derived from the ‘wet’ image have a strong empirical
relationship (r2 = 0.765, p < 0.01) with the clay content of
the sample sites measured (Eq. (3)).
%Clay ¼ exp½0:995 þ 4:13ðMw Þ
ð3Þ
where Mw is the mud abundance estimate from the ‘wet’
image.
Fig. 6. Profile of microphytobenthos and mud abundance from the saltmarsh edge to the main channel, illustrating the attenuating effects of microphytobenthos
on the ‘dry’ 1997 mud abundance image.
M.P. Rainey et al. / Remote Sensing of Environment 86 (2003) 480–490
Table 1
Variation in surface percentage clay measurements made within 2 m2
quadrats in the exposed intertidal sediments in the Ribble Estuary
Quadrat
Coefficient of variation (%)
Quadrat 1 ‘‘mud’’
Quadrat 2 silt
Quadrat 3 sand
20.59
24.71
53.01
Examination of the associated RMS error image revealed
that the unmixing error increased in the fine-grained areas
where the intertidal sediments are covered by microphytobenthos. The mud abundance values are also found to
decrease significantly in the areas of high microphytobenthos abundance. Consequently, it was determined that the
microphytobenthos present attenuated the spectral signal of
the mud resulting in underestimation of mud abundance, as
observed in the ‘dry’ image results. To correct for this
effect, the microphytobenthos and mud abundance estimates
were combined and correlated against actual clay content
(Eq. (4)).
%Clay ¼ exp½1:02 þ 3:55Mw þ 4:29Bw ð4Þ
where Bw is the microphytobenthos abundance estimate
from the ‘wet’ image.
This marginally improved the relationship between the
combined estimates and actual clay content, supporting the
theory that microphytobenthos partially masks the spectral
signal of the mud (r2 = 0.790, p < 0.01). Finally, as in the
‘dry’ image results, pools of surface water severely
attenuated the spectral signal of mud and the channel
walls were misclassified as being areas of high mud
abundance.
487
4.6. Intertidal sand distribution
4.6.1. ‘Dry’ ATM image
A significant, though weak, linear relationship was found
between the sand content of the sample sites and the abundance image estimates (r2 = 0.60, p < 0.01). The associated
RMS error image demonstrated that the unmixing error
within the sand-dominated areas is approximately 1 order
of magnitude greater than that within the mud-dominated
areas of the imagery. Spectral profiles within the sand
abundance image reveal that the estimated sand increases
towards the main channel from the saltmarsh, as expected
from field investigations (Fig. 9). However, the increase in
sand abundance is gradual unlike the steep increase followed
by a plateau observed in reality. An explanation for this
difference is the equally steep increase in moisture abundance
estimates of the sediments at the base of the sloped mudflats.
Consequently, it is proposed that the greater spectral attenuation caused by the increasing interstitial moisture contents
results in an underestimation of sand abundance, which
smoothes the expected steep increase. To correct this, the
moisture abundance image was combined with the sand
abundance image, and this multiple linear regression improved the relationship between the estimated and observed
sand contents significantly (r2 = 0.82, p < 0.01). A further
improvement was gained by adding the microphytobenthos
abundance into the multiple regression (r2 = 0.844, p < 0.01),
Eq. (5). Only small isolated areas of polygonated mud close
to the saltmarsh were misclassified as having high sand
abundance; it is suggested that this is a product of halite
efflorescence in these areas (Bryant et al., 1996).
%Sand ¼ 35:6 33:7Sd þ 59:2Wd 92:7Bd
ð5Þ
where Sd is the sand abundance estimate and Wd is the
moisture abundance estimate from the ‘dry’ image.
Fig. 8. Profile of mud and moisture abundance from the saltmarsh edge to the main channel, illustrating the attenuating effects of surface moisture on the ‘dry’
1997 mud abundance image.
488
M.P. Rainey et al. / Remote Sensing of Environment 86 (2003) 480–490
Fig. 9. Profile of sand and moisture abundance from the saltmarsh edge to the main channel, illustrating the attenuating effects of moisture on the ‘dry’ 1997
sand abundance image.
Fig. 10 shows the relationship between the modelled
prediction (Eq. (5)) and the ground reference sample points.
The disproportionate number of samples with high sand
content reflects the sediment characteristics of the sites
selected, making the data difficult to normalise. However,
the relative linearity of calibration, when moisture and
microphytobenthos is taken into consideration is supported
by laboratory based spectroradiometry measurements
reported in Rainey et al. (2000). As with clay content,
Fig. 10 shows that following the recombination of the
‘unconstrained’ spectral unmixing end-member abundances,
the final image calibration does not result in sand abundances out with the physical limits of 0% to 100% sand
content.
4.6.2. ‘Wet’ ATM image
The sand abundance estimates, produced from the subset
of the ‘wet’ ATM image, demonstrate no significant association with the percentage sand of the in situ samples
(r2 = 0.059). Spectral profiles through the sand abundance
image suggested that high sediment moisture contents were
attenuating the spectral signal of the sand, as in the abundance
imagery of the ‘dry’ image. To correct for this, the sand,
moisture and microphytobenthos abundance estimates were
combined and the result was compared to the actual sand
contents. A strong relationship was evident (r2 = 0.80,
p < 0.01), demonstrating that where high interstitial moisture
contents exist the image analysis technique fails to unmix the
spectral characteristics of sand. As observed in the ‘dry’
abundance imagery, polygonated mud is misclassified in the
sand image as areas of high abundance.
Fig. 10. Comparison between image derived estimate of sand content (Eq.
(5)) and sand content derived from sediment samples (r2 = 0.82, p < 0.01).
Fig. 11. Subset of the percentage clay map produced from the ‘dry’ 1997
abundance imagery.
M.P. Rainey et al. / Remote Sensing of Environment 86 (2003) 480–490
4.7. Clay maps of the Ribble Estuary
The principle outcome of this study, is a complete map of
clay distribution within the intertidal sediments of the Ribble
Estuary. This map is based primarily on the mud and
microphytobenthos abundance images produced from the
‘dry’ ATM imagery (Fig. 11). This is due to the strong
association that exists between the percentage clay content of
the sample sites and the combined mud and microphytobenthos abundance estimates (r2 = 0.815, p < 0.01). The exponential regression equation that defines this relationship was
used to convert the abundance imagery into a map of
percentage clay (Eq. (2)).
5. Discussion
This study has shown that ATM imagery, collected under
relatively well-defined environmental conditions, can be
used to quantitatively map the distributions of sediments
within the intertidal reaches of an estuary. It is suggested
that the distribution of the fine-grained sediments within the
Ribble Estuary on May 30, 1997 is represented in the mud
abundance image produced from the ‘dry’ ATM image (Fig.
4). This proposal is based on the strong relationship
observed between the actual clay content and the estimated
mud abundance of the sample sites (r2 = 0.79). This supports the linear unmixing of ATM imagery to produce
‘mud’ abundance images, as a means of determining the
percentage clay content distribution within the intertidal
areas.
The distribution of mud in the ‘wet’ abundance image
(Fig. 4), is closely correlated with that of the ‘dry’ image
although the latter is marginally more accurate. This suggests
that during the first hour and a half of intertidal exposure,
before the collection of the first line of imagery, the sediments lost sufficient moisture to allow the mud distribution
to be accurately mapped. The topographic relief of the
mudflat areas close to the saltmarsh probably contributed
substantially to the loss of moisture in these regions. However, in the more recently exposed flat sand-dominated areas
the sediments have less opportunity to lose moisture within
the first 1 1/2 h. This may explain the relatively high
moisture abundance estimates of these areas as derived from
the ‘wet’ imagery. It seems reasonable to suggest that the
higher moisture content contributed to the marginally weaker
relationship between the sampled clay content and the mud
abundance estimates derived from the ‘wet’ image, in
comparison to that of the ‘dry’ imagery. This has important
implications on the timing of image acquisition for quantitative mapping of sediment grain size distributions, as also
indicated in Rainey et al. (2000).
The ability to estimate the clay content imagery is
improved when the microphytobenthos abundance is included in the regression equation (r2 = 0.815; Fig. 7). This
suggests that the microphytobenthos is causing an underes-
489
timation of the mud present, as suggested by some abundance profiles (Fig. 5). The degree of underestimation may
well be a function of microphytobenthos biomass. Again, it
seems reasonable to compensate for the presence of microphytobenthos, and this is supported by the improvement of
the relationship between the image estimates and the clay
content of the sample sites. In addition to microphytobenthos, surface water attenuates the spectral signal of the
intertidal sediments to such a degree that they are almost
completely masked out (Fig. 6). Consequently, the mud
abundance estimates drop sharply where surface water is
present. If more detailed in situ knowledge of the surface
pools had been available it may have been possible to
identify all the surface water directly from the water abundance image. This may then have allowed for a suitable
correction to be applied to the affected areas of the abundance imagery. However, this further justifies the need to
maximise the intertidal drying time to minimise the affect of
surface water.
The clay abundance image, produced from the ‘dry’ mud
abundance image, is an accurate representation of clay
distribution within the Ribble Estuary, for May 30th 1997.
Excellent correlations between the image abundance estimates and the sample clay content were found. It is
interesting to note that much of the remaining scatter in
these relationships may be accounted for by sampling and
analytical error. The methodology presented therefore forms
a useful protocol for mapping intertidal sediment grain size
distributions using multispectral sensors that measure within
the SWIR from airborne platforms.
The ability to map sand abundance from ATM imagery is
also dependent on the moisture content and microphytobenthos cover of the sediments. This is demonstrated by the
contrasting relationships of the ‘wet’ and ‘dry’ sand abundance imagery, with the sand content of the sample sites.
During the 3 h of sediment exposure between the acquisition
of these images, it is suggested that the sediments were able
to lose sufficient interstitial moisture to reduce significantly
the attenuation of sand’s spectral characteristics in particular. This resulted in the map of sand distribution produced
from the ‘dry’ image being considerably more accurate
(r2 = 0.60) than that produced from the ‘wet’ imagery
(r2 = 0.059). It has also been shown that by combining the
water and microphytobenthos abundance image with the
sand abundance image, the ability to estimate percentage
sand is improved considerably (r2 = 0.844 for ‘dry’ image
data; r2 = 0.8 for ‘wet’ image data).
Given the exceptional atmospheric clarity during image
acquisition and application of a dark pixel subtraction to
the image, subsequent use of the calibrations described
here may be achieved on ATM imagery of the Ribble
which has been normalised to the original May 1997
image data set. More generic application of these calibrations will be susceptible to atmospheric correction
techniques employed and would need to be re-evaluated
appropriately.
490
M.P. Rainey et al. / Remote Sensing of Environment 86 (2003) 480–490
6. Conclusions
An image analysis technique for mapping percentage clay
content distribution within the intertidal zones of an estuary
at low tide has been presented. This technique was proven
effective by the production of accurate maps of intertidal
sediment abundance from raw ATM imagery. It has been
shown that the period of sediment exposure before image
collection is important to the accuracy of the final results,
supporting the findings of earlier reflectance spectroradiometry. Although accurate maps of clay abundance may be
created from imagery acquired following a short period (1 1/
2 h) of sediment exposure, a more accurate product is
produced from imagery following an extended period (4 1/
2 h) of exposure. This determinant has been shown to have
even more serious implications on the ability to map sand
contents within this region. However, by combining image
end member abundances through multivariate regression
analysis, it is possible to produce suitably accurate calibration models for mapping intertidal grain size distributions.
Acknowledgements
This research was funded the by the University of Stirling
and BNFL through Westlakes Scientific Consulting. The
authors would like to thank the Natural Environment Research Council for the provision of the airborne remote
sensing data. The authors would also like to acknowledge Bill
Jamieson and David Aitchison who drafted a number of figures and numerous colleagues at Westlakes Scientific Consulting and the University of Stirling for aiding with reference
data collection across the whole of the Ribble Estuary.
References
Adams, J. B., Smith, M. O., & Gillespie, A. R. (1993). Imaging spectroscopy: Interpretation based on spectral mixture analysis. In C. M.
Pieters, & P. Englert (Eds.), Remote Geochemical Analysis: Elemental
and Mineralogical Composition, vol. 7 ( pp. 145 – 166). Cambridge
Univ. Press.
Black, K. S., & Paterson, D. M. (1997). Measurement of the erosion
potential of cohesive marine sediments: A review of current in situ
technology. Journal of Marine Environmental Engineering, 26, 43 – 83.
Brown, J. (1997). The behaviour of radionuclides in the Ribble Estuary,
NW England. Unpublished PhD Thesis, University of Reading, UK.
Bryant, R., Tyler, A. N., Gilvear, D. J., McDonald, P., Teasdale, I., Brown,
J., & Ferrier, G. (1996). A preliminary investigation into the VNIR
spectral characteristics of inter-tidal sediments. International Journal
of Remote Sensing, 17, 405 – 412.
Cook, G. T., MacKenzie, A. B., McDonald, P., & Jones, S. J. (1997).
Remobilization of Sellafield-derived radionuclides and transport from
the NE Irish Sea. Journal of Environmental Radioactivity, 35, 227 – 241.
Donoghue, D. N. M., Reid Thomas, D. C., & Zong, Y. (1994). Mapping
and monitoring the intertidal zone of the east coast using remote sensing
techniques and a coastal monitoring GIS. Marine Technology Society,
28, 19 – 29.
Green, A. A., Berman, M., Switzer, P., & Graig, M. D. (1988). A transformation for ordering multispectral data in terms of image quality with
implications for noise removal. IEEE Transactions on Geoscience and
Remote Sensing, 26, 65 – 74.
HMIP (1994). A survey of radioactivity in the Ribble Estuary: Activity
concentrations and estuarine dynamics. Her Majesty’s Inspectorate of
Pollution. Environment Series Report No. 1.
Mackenzie, A. B., Cook, G. T., McDonald, P., & Jones, S. R. (1998). The
influence of mixing timescales and re-dissolution processes on the distribution of radionuclides in northeast Irish Sea sediments. Journal of
Environmental Radioactivity, 39, 35 – 53.
Mertes, L. A. K., Smith, M. O., & Adams, J. B. (1993). Estimating suspended sediment concentrations in surface waters of the Amazon River
Wetlands from Landsat Images. Remote Sensing of Environment, 43,
281 – 301.
Painter, T. H., Roberts, D. A., Green, R. O., & Dozier, J. (1998). The effect
of grain size on spectral mixture analysis of snow-covered area from
AVIRIS Data. Remote Sensing of Environment, 65, 320 – 332.
Paterson, D. M. (1997). Biological mediation of sediment erodibility: Ecology and physical dynamics. In N. Burt, R. Parker, & J. Watts (Eds.),
Cohesive sediments ( pp. 215 – 229). Wiley.
Rainey, M. P. (1999). Airborne remote sensing if estuarine intertidal radionuclide concentrations. PhD thesis, University of Stirling: UK.
Rainey, M. P., Tyler, A. N., Bryant, R., Gilvear, D. J., & McDonald, P.
(2000). The influence of surface and interstitial moisture on the spectral
characteristics of intertidal sediments: Implications for airborne image
collection and processing. International Journal of Remote Sensing, 21,
3025 – 3038.
Settle, J. J., & Drake, N. A. (1993). Linear mixing and the estimation of
ground cover proportions. International Journal of Remote Sensing, 14,
1159 – 1177.
Thomson, A. G., Eastwood, J. A., Yates, M. G., Fuller, R. M., Wadsworth,
R. A., & Cox, R. (1998). Airborne remote sensing of intertidal biotopes:
BIOTA. Marine Pollution Bulletin, 37, 164 – 172.
Thomson, A. G., Fuller, R. M., Sparks, T. H., Yates, M. G., & Eastwood, J.
(1998). Ground and airborne radiometry over intertidal surfaces: Waveband selection for cover classification. International Journal of Remote
Sensing, 19, 1189 – 1205.
Tyler, A. N. (1999). Monitoring anthropogenic radioactivity in salt marsh
environments through in situ gamma ray spectrometry. Journal of Environmental Radioactivity, 45, 235 – 252.
Tyler, A. N., Sanderson, D. C. W., Scott, E. M., & Allyson, J. D. (1996).
Investigations of spatial variability and fields of view in environmental
gamma ray spectrometry. Journal of Environmental Radioactivity, 33,
213 – 235.
Van der Wal, D., Pye, K., & Neal, A. (2002). Long-term morphological
change in the Ribble Estuary, northwest England. Marine Geology, 189,
249 – 266.
Wilson, A. K., Mockeridge, W., & White, S. (1996). The NERC airborne
remote sensing facility, development and implementation of the integrated data system. Proceedings of the 22nd Annual Conference of the
Remote Sensing Society, Durham, 11 – 14 September ( pp. 263 – 270).
Yates, M. G., Jones, A. R., McGrorty, S., & Goss-Custard, J. D. (1993).
The use of satellite imagery to determine the distribution of intertidal
surface sediments of the Wash, England. Estuarine Coastal and Shelf
Science, 36, 333 – 334.