Académique Documents
Professionnel Documents
Culture Documents
ABSTRACT
Coastal saltmarsh and their constituent components and processes are of an interest scientifically due to their ecological
function and services. However, heterogeneity and seasonal dynamic of the coastal wetland system makes it challenging
to map saltmarshes with remotely sensed data. This study selected four important saltmarsh species Pragmitis australis,
Sporobolus virginicus, Ficiona nodosa and Schoeloplectus sp. as well as a Mangrove and Pine tree species, Avecinia and
Casuarina sp respectively. High Spatial Resolution Worldview-2 data and Coarse Spatial resolution Landsat 8 imagery
were selected in this study. Among the selected vegetation types some patches ware fragmented and close to the spatial
resolution of Worldview-2 data while and some patch were larger than the 30 meter resolution of Landsat 8 data. This
study aims to test the effectiveness of different classifier for the imagery with various spatial and spectral resolutions.
Three different classification algorithm, Maximum Likelihood Classifier (MLC), Support Vector Machine (SVM) and
Artificial Neural Network (ANN) were tested and compared with their mapping accuracy of the results derived from
both satellite imagery. For Worldview-2 data SVM was giving the higher overall accuracy (92.12%, kappa =0.90)
followed by ANN (90.82%, Kappa 0.89) and MLC (90.55%, kappa = 0.88). For Landsat 8 data, MLC (82.04%) showed
the highest classification accuracy comparing to SVM (77.31%) and ANN (75.23%). The producer accuracy of the
classification results were also presented in the paper.
.
Keywords: Saltmarsh, Worldview-2, Landsat 8 OLI, Classification, Assessment, MLC, SVM, ANN.
1. INTRODUCTION
Saltmarsh is an intertidal community and treated as Ecological Endangered Community (EEC) in Australia [1]. For
these reasons, monitoring and dynamic change analysis of saltmarsh is a pressing issue and scientists are much more
dependent on high quality remote sensing data for mapping and monitoring of saltmarsh and their proactive management
[2-4]. However, this community usually lives with dominant mangroves. This causes a problem in selection of scale and
radiometric properties of the remotely sensed data when performing classifications. The scale or spatial resolution is the
projected area on the ground associated with the radiance measurement of the sensors. It is the smallest distinguishable
spatial unit (ground resolution element) recorded in a remotely sensed data. For example, the spatial resolution is 2 m for
Worldview-2 data and 30m for Landsat 8 OLI or EO-1 Hyperion data. Sometimes the spatial resolution denotes the
ground sampling distance in an image after image re-sampling that can be varied from the actual spatial resolution
recoded by the sensor. Spectral resolution means the range over the electromagnetic spectrum the energy is measured
and recorded by the sensor. The use of higher spectral resolution images usually improve the capability of detecting
spectral variability within a land cover by increasing the discrimination capacity [5]. But spatial resolution may be more
important than spectral resolution when spatial extent of land cover types is an issue. However, both spatial and spectral
resolution are two fundamental characteristics of a remotely sensed image for any application, like as classification or
feature extraction.
Remote sensing image classification accuracy is affected by two factors: firstly, influence of boundary pixels and
secondly, finer spatial resolution that increases the spectral radiometric variation of land cover types [6]. Optimum
spectral and spatial resolution determination for vegetation mapping has been an ongoing area of research in remote
sensing [7]. It becomes more challenging when dominant and fragmented plant species patches are distributed together
in a same community, similarity to our study site. Our study area supports an extensive temperate saline coastal wetland
system comprising mangrove, saltmarsh and mixed mangrovesaltmarsh habitats [8]. About 4,257 ha of these wetlands
Sensing for Agriculture and Food Quality and Safety VIII, edited by Moon S. Kim, Kuanglin Chao, Bryan A. Chin,
Proc. of SPIE Vol. 9864, 986409 2016 SPIE CCC code: 0277-786X/16/$18 doi: 10.1117/12.2222960
Study site
The study area is located (figure 1 ) in Tomago, Australia which is approximately 10 km north of Newcastle (NSW)
(Figure 1). The topography is generally flat, low lying and subject to periodical flooding. A series of drainage channels
and levee banks dissect the study area from sea.
Input Data
Two different platform satellite imagery of Worldview-2 and Landsat 8 OLI were used for this research. Worldview-2
has a pixel resolution of 0.46m in the panchromatic mode and 1.84 m resolution in the multispectral mode whereas
recently launched Landsat 8 OLI consist of nine spectral bands with a spatial resolution of 30 meters for Bands 1 to 7
and 9. New band 1 (ultra-blue) is useful for coastal and aerosol studies. Another new band 9 is useful for cirrus cloud
detection. The resolution of Band 8 (panchromatic) is 15 meters. Thermal bands 10 and 11 are useful in providing more
accurate surface temperatures and are collected at 100 meters [12]. Approximate scene size is 170 km north-south by 183
km east-west (106 mi by 114 mi). In our study we did not use thermal and cirrus bands (table 1)
380000 384000
Figure 1. Location of the study area on World Imagery (Google earth map) (Left) and Landsat 8 RGB (Right).
i
n
8
2
-0.113
0.1 5
PC Iknd 1 (AILLarraEL1111.50:PcA-s.trf
-a:os
PC Bond 2 r.C43.tif
'0i.;,13_8',.;,),' a.@ ois
Figure 2: Principal Components Analysis (PCA) based endmember selection for classification of Landsat 8 OLI data.
Scatterplot of the three PCs from Landsat OLI data to select seven endmember: (Clockwise) PCA 1vs PCA2 : Mangrove,
Grass, Phragmitis, Water; PCA 1 vs PCA3: Casuarina, Sporobolus, Water ; PCA 2 vs PCA3: Phragmitis.
Selected classifier
Maximum Likelihood Classifier (MLC), Support Vector Machine (SVM) and Artificial Neural Network (ANN) have
been selected for classification purposes. MLC is one of the statistical classifiers. It is based on that depends on the idea
that the distribution of the data in each class are normally distributed and called as parametric classifier. While the SVMs
identify the optimal hyperplane that creates the largest distance, or margin, between the vectors for the two classes. In
situations where the training sets for the classes are not linearly separable, the classification is performed in the mapped
space (higher dimensional space) instead of the input feature space [18]. ANN works based on the processing node that
corresponds to the neuron of the human brain where each node receives and sums a set of input values and passes this
sum through an activation function providing the output value of the node [19]. Here input layers represents for spectral
bands of the data and output layers are the number of classes for the network.
Accuracy Assessment and kappa analysis
Accuracy assessment results were discussed based on the confusion error matrix, overall map accuracy and kappa values.
For individual class, producer and user accuracy were computed based on the dominant class in each reference plot [20-
22].
Mapping Accuracy (MA)
MA [23] was computed based on the following formula
3. RESULTS
Principal Component Analysis
The PCA eigenvalue showed that the first three PCA components accounted for almost 99.50% of the total variance for
Worlview-2 data and 99.76% for of the total variance for Landsat 8 OLI data. However, if we consider the pixel
resolution of Worldview-2 data then rest 0.50% variance is really important for 2X2 m pixels. But we can ignore rest
0.25% variance (Table 2) based on 30x30 m pixels resolution for Landsat OLI data to select independent endmember
dataset for classification. In this perspective our calibration data from Landsat OLI extracted maximum number of
endmember (for Landsat 8 it is 7) from uncorrelated PCA bands. Existing studies suggest that the pixels located on the
corner of the scatterplot (Figure 2) can be treated as endmember [24].
When the spectral curve of these selected endmember were compared with the original image spectra, it was also proved
that maximum of seven endmember (Figure 3) could be extracted from Landsat OLI data. Based spectral curve error
bar it is very clear that all saltmarsh species are overlapped with each other within all bands of the spectrum. But
Mangrove and Casuarina tree species are clearly separable from each other up to bands 5. Similarly, band 1-4 are clear
indication to separate Marshy wetland from water.
Spectral profile Analysis:
On the other hand for Worldview-2 data (Figure 4), reflectance properties of water in bands 6-8 (Red Edge, NIR1 and
NIR 2) and reflectance of marshy wetland in band 7 are significantly different from other classes.
0.4
0.35
Mangrove
0.3 Casuarina
0.25 Phragmitis
% Reflectance
Sporobolus
0.2 PerrenialGrass
0.15 Water
MarshyWetland
0.1
0.05
0
1 2 3 4 5 6 7
-0.05
Different bands of Landsat 8 OLI data
Figure 3: Different endmember (class) extracted from 7 bands of Landsat 8 OLI data
0.7
0.6
0.5
t Mangrove
a) 0.4 Casuarina
"
C -II- Water
co
0.3 tPhragmitis
T.5
a) --Sporobolus
-411- MarshyWetland
= 0.2 PerrenialGrass
0.1
0
1 2 3 4 5 6 7 8
-0.1
Different bands of Worldview 2 data
Another sources of classification error comes from the similarity of spectral profile of Sporobolus virginicus with
Schoenoplectus sp. similar trend was also observed in other two classifier, SVM and ANN Moreover Schoenoplectus
was the most difficult classes for which to collect ground reference data, due to relatively small size of the parcels. Based
on these three classifier, SVM showed the maximum overall accuracy among these three.
To test individual species accuracy, Mapping Accuracy (MA %) result is shown on table 7 to compare the result from
two different platform sensors.
4. DISCUSSION
Our finding from parametric (MLC) with non-parametric classifiers (SVM and ANN) in Worldview-2 data are
comparable to the results given the previous studies. For example, [25] found a better performance of using SVM when
it was compared against ANN using SPOT-5. One of the limitation of our finding with SVM is 0 with Schoeloplectus sp
where MLC provided the better accuracy to delineate the narrow patch and some fragmented location of this species.
This is might be due to the training sample size. Based on [26] classification accuracy is affected by training sample size