A Hierarchical Multi-classifier Framework for Landform Segmentation Using Multi-spectral Satellite Images - A Case Study over the Indian Subcontinent

TL;DR: A hierarchical method for landform classification for identifying a wide variety of landforms occurring over parts of the Indian subcontinent is proposed and the results are compared with two other methods of classification.

Abstract: There is an increasing need for automatically segmenting the regions of different landforms from a multispectral satellite image. The problem of Landform classification using data only from a 3-band optical sensor (IRS-series), in the absence of DEM (Digital Elevation Model) data, is complex due to overlapping and confusing spectral reflectance from several different landform classes. We propose a hierarchical method for landform classification for identifying a wide variety of landforms occurring over parts of the Indian subcontinent. At the first stage, the image is classified into one of three broad categories: Desertic, Coastal or Fluvial, using decision fusion of three SVMs (Support Vector Machine). In the second stage, the image is then segmented into different regions of landforms, specifically belonging to the class (category) identified at stage 1. To show the improvement in accuracy of our classification method, the results are compared with two other methods of classification.

Summary

Introduction

• Landform classification had been attempted in the past either using data from SAR (Synthetic Aperture Radar) or DEM or by integrating data form multiple sensors.
• Topographic attributes were calculated from DEM and classified using SOM.
• An accuracy of 97% was also reported by iterative selection of point sample training set.
• These has been prime motivation of their work.

II. PROPOSED FRAMEWORK

• Based on interactions with GIS and Geomorphological experts, the authors came to understand that the data samples were acquired from three (3) major categories of areas/zones: Desertic, Coastal and Fluvial.
• The following landform classes are being considered for the purpose of landform identification from satellite images of the coastal belt: Creek, Forested Swamp, Sandy Beach, Coastal Bars, Sea, Flood Plains and Alluvial Plains.
• The authors propose a hierarchical method of landform classification (Fig. 1) that performs super-group classification at the first level i.e. it determines the supergroups of the input image.
• This enables us to search for the probable set of landforms occurring in the input image, only under the particular super-group that has been determined at the first step of processing.
• The nodes used to label the different processing methods, as given in Fig. 1 are: DP-Desertic Processing, CP-Coastal Processing and FP-Fluvial Processing methods.

A. Super-Group Classification

• This is the topmost stage of the proposed hierarchical classification as shown in Fig.
• Thus the classification at this stage is primarily based on the multi-spectral intensities features, using SVM as a classifier.
• The number of samples used for training and testing, and the accuracies of classification obtained are given in Table I. Sum rule [9] has been used for fusing the decisions, as it has been shown to work better in [10].
• Table II shows the number of testing samples misclassified during testing phase before fusion, by each of the three classifiers (SVM-D, SVM-C and SVM-F).
• All the SVM classifiers are used with a polynomial kernel (of degree 2).

B. Sub-Group Classification

• This stage consists of a set of processes which are detailed at the bottom part of the flowchart in Fig. 1, following super-group classification.
• 1) Processing Modules for the Desertic Type of Landforms:.
• The steps of processing for identification of landform in coastal images are as follows: CP-I segregates the water-bodies from land by thresholding the intensity of blue color (in NIR, R and G band).
• FP II B classifies channel, plain and ox-bow. FP-III A (Connected Component Labeling & Adjacency Information) identifies Bars, Flood Plain, and Alluvial Plain.

III. RESULTS AND DISCUSSION

• Training and testing samples from the landform data were acquired with the help of hand-labeled data to improve the results of sub-classification stages.
• In other cases this was not estimated, as the authors had used trivial image processing methods (Template Matching, Connected Component Labeling, Adjacency Information, etc.) and not classifiers for segmentation and labeling of the pixels.
• The overall accuracy of the proposed method (quantitatively measured for a few landforms and visually compared in other cases using results on raster images) is better due to the hierarchical organization of a set of classifiers.
• Each classifier in their proposed framework, solves a specific part of the overall problem of classification, for a small set (2-4) of classes within a limited domain.
• Though their method takes more time than the method proposed by Gagrani et al. [13], the performance of their method is superior (see Fig. 2-4).

Figures

Figure 1: The proposed hierarchical classification scheme.

Figure 2: (a) Input Desertic image; b) Hand-Labeled Data; Output: (c) Unsupervised method [12] & (d) Method proposed in [13]; (e) Our proposed method.

Figure 4: (a) Input Fluvial image; b) Hand-Labeled Data; Output: (c) Unsupervised method [12] & (d) Method proposed in [13]; (e) Our proposed method.

Figure 3: (a) Input Coastal image; b) Hand-Labeled Data; Output: (c) Unsupervised method [12] & (d) Method proposed in [13]; (e) Our proposed method.

Full text

A Hierarchical multi-classifier Framework for Landform Segmentation using
Multi-spectral Satellite Images - A case study over the Indian subcontinent
Utthara Gosa Mangai, Suranjana Samanta, Sukhendu Das
VP Lab, Dept. of CSE
IIT Madras, Chennai, India
utthara@cse.iitm.ac.in, ssamanta@cse.iitm.ac.in, sdas@iitm.ac.in
Pinaki Roy Chowdhury
Defence Terrain Research Laboratory (DRDO)
Delhi, India
rcpinaki@yahoo.com
Koshy Varghese
Department of Civil Engineering
IIT Madras, Chennai, India
koshy@iitm.ac.in
Manisha Kalra
Parc Vista Condo, 460 Corporation Road, #03-08
Singapore, 649815
manishakalra007@gmail.com
Abstract—There is an increasing need for automatically
segmenting the regions of different landforms from a multi-
spectral satellite image. The problem of Landform classification
using data only from a 3-band optical sensor (IRS-series), in the
absence of DEM (Digital Elevation Model) data, is complex due
to overlapping and confusing spectral reflectance from several
different landform classes. We propose a hierarchical method
for landform classification for identifying a wide variety of
landforms occurring over parts of the Indian subcontinent. At
the first stage, the image is classified into one of three broad
categories: Desertic, Coastal or Fluvial, using decision fusion of
three SVMs (Support Vector Machine). In the second stage, the
image is then segmented into different regions of landforms,
specifically belonging to the class (category) identified at stage
1. To show the improvement in accuracy of our classification
method, the results are compared with two other methods of
classification.
Keywords-Landform Classification; Support Vector Machine;
Hierarchical Classification; Decision Fusion;
I. INTRODUCTION
Landform classification is a problem of identifying a set
of predefined categories of landforms. Probabilistic clas-
sification of landforms is significant and a hard problem
because of the versatility and variations in the categories
of landforms, their patterns, features, association rules and
signatures. However, there exists huge scope of application
of such an area of work, in remote sensing and GIS (Geo-
graphic Information System), trafficability and surveillance
in military domain, cartographical updates and geological
surveys etc. We pose the problem as: Given a satellite image
of a large area, identify the different landform segments
from the image, with a confidence measure for each detected
landform segments.
Indian Landform image data and the corresponding hand-
labeled data as ground truth were provided to us by GIS ex-
perts. Landform classification had been attempted in the past
either using data from SAR (Synthetic Aperture Radar) or
DEM or by integrating data form multiple sensors. The ob-
jective of our research was to classify the Indian landforms
from a single sensor (3-band) image. This objective led
to the development of a hierarchical classification scheme.
Since different landforms have different signatures, a single
classifier will not be sufficient for classification. Hence,
a hierarchical landform classification method is proposed,
where separate sets of training samples are used for separate
classifiers at different levels of the hierarchical framework.
Related works have been done by [1], [2], [3], [4], [5],
[6] to solve the problem mostly using DEM data. The
methods used for landform classification can be broadly
divided into three different categories: (i) Use of a statistical
(Bayes, Fuzzy k-means, Maximum Likelihood etc.) criteria;
(ii) Set of rules in a framework for multi-level threshold-
ing (with different parameters); and (iii) Neural classifier-
Self-organising map (SOM), feed forward neural network
(FFNN) etc. In all of these cases, results have not been
shown in the absence of DEM, which seems to provide
a strong clue for visual experts to detect and discriminate
closely lying neighboring landforms. The main problem with
the use of DEM is its free availability at a higher resolution
(below 10 or 25 m, say) and online accessibility for a
particular region based on instantaneous demand.
In [3], a numerical method for classification and character-
ization of landforms on Mars was proposed. Topographic
attributes were calculated from DEM and classified using
SOM. Landform of 21 classes were broadly classified into
five categories (Highlands, craters, lowlands, high-relief and
channels). Reference [7] reports results of employing a
fuzzy c-means classification for a DEM data. Remotely
sensed topographic data gathered by orbiting satellites were
transformed into semantically meaningful maps of landforms
[8]. The mapping is achieved by means of scene seg-
2010 Fourth Pacific-Rim Symposium on Image and Video Technology
978-0-7695-4285-0/10 $26.00 © 2010 IEEE
DOI 10.1109/PSIVT.2010.58
306

mentation using K-means-based agglomerative segmentation
and watershed-based segmentation followed by supervised
classification of segments. Hosokawat and Hoshi [1] used
SOM to classify landforms into hill, plateau, fan based on a
land cover map and DEM and reported an accuracy of 77%
using four landform samples. Reference [4] used data from
multiple sensors and proposed an approach which employs
a class dependent feature selection in conjunction with pair-
wise Bayesian classifiers and reports an accuracy of 96%.
Hengl and Rossiter [5] used maximum-likelihood classifier
to classify landforms into 21 legends. Authors used DEM
data to extract nine terrain parameters. An accuracy of 97%
was also reported by iterative selection of point sample
training set.
In spite of the vastness of the work done on landform
classification, it can be seen that all of these require either
DEM (in most cases) or SAR/ LIDAR data. For certain
online applications the DEM data may not be available
(as it is not a direct output of a sensor). It remains a
challenging task and a hard problem to perform landform
classification using an optical sensor image alone as input.
This is the main reason why researchers have not been done
to solve this problem with a reasonable degree of accuracy
using only a multi-band optical image. These has been
prime motivation of our work. We designed a hierarchical
multi-classifier framework to solve this problem. Section
II describes the overall framework and various stages in
Hierarchical Landform Classification. Results are discussed
in Section III. Section IV presents the conclusion of the
paper.
II. P
ROPOSED FRAMEWORK
Based on interactions with GIS and Geomorphological
experts, we came to understand that the data samples were
acquired from three (3) major categories of areas/zones:
Desertic, Coastal and Fluvial. We hence used a divide
and conquer approach to design a hierarchical system of
classification. This made the problem tractable, as about
30 different classes of landforms were available for seg-
mentation, and the features (spectral) were overlapping and
confusing over many classes of landforms. In the following,
we first discuss the three major categories of landform
classes followed by the design of our proposed framework.
1) Fluvial landforms: These consist of the landforms that
are produced by the action of stream or river. The
following landforms classes are being considered for
the purpose of landform identification from satellite
images of fluvial areas: (a) Active - Active Channel,
Bars, Flood Plains, Alluvial Plains, (b) Relict (Dry)
-Oxbow lakes, Forested swamps, Alluvial Plains.
2) Coastal landforms: The following landform classes
are being considered for the purpose of landform
identification from satellite images of the coastal belt:
Creek, Forested Swamp, Sandy Beach, Coastal Bars,
Sea, Flood Plains and Alluvial Plains.
3) Desertic landforms: The following classes of land-
forms have been considered for detection from satellite
images taken from desert areas: Barchan Dunes, Lon-
gitudinal Dunes, Transverse Dunes, Parabolic Dunes,
Barchanoid, Sandy Plains, Rocky Exposures, Insel-
berg, Remnant Stony Surfaces (Salt Flats) and Playa.
The proposed hierarchical classification scheme is shown
in Fig. 1. We assume that an input image given to the
landform extraction system will not contain different types
of landforms belonging to any two of the three different
categories: Desertic, Coastal and Fluvial (termed as the
Super-Group classes). We propose a hierarchical method of
landform classification (Fig. 1) that performs super-group
classification at the first level i.e. it determines the super-
groups of the input image. The uncertainty of classification
into different landform zones is the largest at this stage and
a near 100% crisp classification is expected from the super-
group classifier in use. This enables us to search for the
probable set of landforms occurring in the input image, only
under the particular super-group that has been determined at
the first step of processing. Thus, the uncertainty involved
in detecting the landforms in an image, is reduced to
approximately one-third of the total number of classes. The
2
nd
stage of processing detects sub-classes of a landform
category (i.e. either Desertic or Coastal or Fluvial) detected
at the super-group stage.
At the leaves of the hierarchical classification tree given in
Fig. 1, are all the landforms of interest organized suitably un-
der their respective processing methodologies. Based on the
output of the super-group classifier, all algorithms tailored
for the extraction of features of each subset of landforms
would be used in parallel for accurate segmentation of
leaf landform nodes. Thus, at the bottommost level of the
hierarchical tree, landform-specific modules are used to
detect landforms accurately. Such modularity in the nature
of landform extraction is designed, keeping in mind the huge
amount of uncertainty, combinations of occurrences and
adjacency of various landforms as well as the subjectivity
in the landform definitions, which are undoubtedly too
confusing for any one-classifier to capture. The nodes used
to label the different processing methods, as given in Fig.
1 are: DP-Desertic Processing, CP-Coastal Processing and
FP-Fluvial Processing methods. These processing methods
are discussed in detail in the following.
A. Super-Group Classification
This is the topmost stage of the proposed hierarchical
classification as shown in Fig. 1. A Support Vector Machine
(SVM) based classification technique has been adopted in
our design for the task of identifying an input image as
belonging to one of the Desertic, Coastal or Fluvial landform
Super-Group categories. In order to capture and exploit
307

Figure 1: The proposed hierarchical classification scheme.
the variability among the different multi-spectral images
belonging to each of the super-groups, histograms of all
the 3 bands, namely NIR, Red and Green were used as
features for classification. Thus the classification at this
stage is primarily based on the multi-spectral intensities
(color) features, using SVM as a classifier. The number of
samples used for training and testing, and the accuracies of
classification obtained are given in Table I. Sum rule [9] has
been used for fusing the decisions, as it has been shown to
work better in [10].
Table II shows the number of testing samples misclassified
during testing phase before fusion, by each of the three
classifiers (SVM-D, SVM-C and SVM-F). SVM-D denotes
the SVM classifier trained with 150 samples of Desertic
landforms and 300 from the remaining two classes. In a
similar way, SVM-F and SVM-C were also trained. All
the SVM classifiers are used with a polynomial kernel
(of degree 2). Number of misclassified samples in table
II of each classifier includes False Rejection and False
acceptance. Accuracy was improved by 1.3% (average) after
using decision fusion technique. Details of the misclassified
samples and the overall accuracy at the super-group stage
of classification, before and after fusion are given below:
Number of misclassified samples before Fusion: 16
Table I: Accuracy Obtained During Testing Phase For Dif-
ferent Landforms.
Landforms
Samples Accuracy in %
Training Testing Before fusion After fusion
Desertic 150 250 99.07 99.47
Fluvial 150 250 99.47 100
Coastal 150 250 99.2 99.6
Table II: Different Classifiers And Their Misclassified Sam-
ples During Testing Before Fusion.
Classifier Samples Misclassified Samples
Training Testing before fusion
SVM-D 450 750 7
SVM-F 450 750 4
SVM-C 450 750 6
Number of misclassified samples after Fusion: 7
Overall Accuracy before fusion: 97.87%
Overall Accuracy after fusion: 99.07%.
308

B. Sub-Group Classification
This stage consists of a set of processes which are detailed
at the bottom part of the flowchart in Fig. 1, following
super-group classification. Based on the experimentations
with landform samples, as well as discussions with geo-
scientists interacting with us, it was observed that we need
to design a hierarchical framework tailored for the extrac-
tion of a particular landform of interest (being analyzed),
because different landforms may require different methods
of processing. In the following, we discuss the processing
methodologies used to obtain results on the landform data
samples supplied to us by experts. The four main features
used for classification are:
• Local mean and/or variance of the approximation sub-
band of the DWT (Discrete Wavelet Transform) for all
the three bands (NIR, R and G). Daubeschies 10-tap
filter is used to extract the DWT coefficients which
represent the texture features in our case.
• Local mean and/or variances of the multi-resolution
color intensities of three bands (NIR, R and G).
• Spatial adjacency of the landforms are stored in the
form of an adjacency matrix, which was formulated
using domain knowledge obtained from GIS experts.
• Connected Component Labeling is used to extract the
shape features which are used to classify certain shapes.
In the following three sub-sections, we present the prop-
erties, features and methods used to analyze three major
categories of landforms discussed in this work.
1) Processing Modules for the Desertic Type of Land-
forms: Two types of processing occur predominantly for
extraction of desertic landforms. These can thus be grouped
together as DP-I and DP-II, each corresponding to methods
based on texture features and shape features, respectively.
The steps of processing for identification of landform in
desertic images are as follows:
• DP-I (multi-class SVM) uses a SVM trained using local
mean from ’Approximation’ of DWT of all three spec-
tral bands (NIR, R and G), for differentiating between
Dunes, Salt flats, Rocky exposure, Barchanoids and
Inselberg.
• DP-II A (Template Matching), further classifies dunes
into parabolic, longitudinal or sandy plain. Templates
for parabolic and longitudinal dunes are obtained from
the training samples. The templates are matched with
the landform image using cross correlation. Output of
this operation is compared with two thresholds (one for
parabolic and another for longitudinal) for labeling as
parabolic or longitudinal dunes. The rest are labeled as
Sandy Plains.
• DP-II B (Area-based operation): Both saltflats and
Playa have similar signatures, it was observed that
Playa was comparatively smaller than salt flats.
2) Processing Modules for the Coastal Type of Land-
forms: It can be observed that intensity-based features have
a major role to play for extraction of coastal landforms.
This is possibly because of the small resolution, fineness and
non-texture information that most of the coastal landforms
have been found to possess. Association rules have also been
employed in order to encode domain-experts knowledge in
observing certain key characteristics of coastal landforms
within the system. The steps of processing for identification
of landform in coastal images are as follows:
• CP-I (Threshold) segregates the water-bodies from land
by thresholding the intensity of blue color (in NIR, R
and G band). If the intensity is very less it is considered
as water-bodies otherwise it is considered as land.
• CP-II A (Adjacency Information) uses adjacency infor-
mation (nearness using an Euclidean measure) to iden-
tify creeks from the water bodies. Creeks are detected
by identifying fine and narrow extents of water bodies
extending into the land. This is done by observing
within a small window whether the extent of sea is
covered by land on the both the sides. This decision is
taken by observing whether the pixels and the border
of the window are non-sea type. The rest of the water
bodies are detected as sea.
• CP-II B (SVM) classifies plain, beach and forested
swamp using a SVM trained using the mean of multi-
resolution color intensity features, computed as:
X
i,j
=[μ(I
n
i,j
) μ(I
r
i,j
) μ(I
g
i,j
)] (1)
where, X
i,j
represents a 3D feature vector correspond-
ing to (i, j)
th
pixel. I
n
i,j
, I
r
i,j
and I
g
i,j
represents in-
tensity values of (i, j)
th
pixel in three spectral bands
(NIR, R and G) of the input image respectively and
μ(h) represents the mean of h computed using windows
of size 5*5, 17*17 and 31*31.
• CP-III A (Connected Component Labeling & Adja-
cency Information)
– Coastal bars possess unique characteristic prop-
erty of being enclosed by sea on all sides. A
connected component labeling algorithm [11] is
employed over pixels whichever is classified as
Plains (detected in CP II B) to determine if all
set of connected pixels are surrounded by sea.
– Among different Plains (detected in CP II B)
whichever is closer to creek, are classified as flood
plains and others are classified as alluvial plain
based on adjacency information.
3) Processing Modules for the Fluvial Type of Landforms:
All the methods that have been employed for detection of
fluvial landforms rely heavily on intensity-based features.
Since fluvial landforms are produced by the action of river
or an active channel, a satellite image taken of a fluvial area
must therefore necessarily contain an active channel within
309

it. The steps of processing for identification of landforms in
fluvial images are as follows:
• FP -I (SVM) uses a multi-class SVM trained using
histogram features of all three spectral bands, is used
to differentiate between active and dry zones. The
channels which were active were blue in color and
others were red in color (based on NIR, R and G bands).
Since the two signatures are different they are trained
and classified separately.
• FP-II A and FP II B (SVM) uses a SVM classifier
trained using mean of color features similar to CP II
B. FP II A classifies channel, forested swamp, plain and
ox-bow. FP II B classifies channel, plain and ox-bow.
• FP-III A (Connected Component Labeling & Adja-
cency Information) identifies Bars, Flood Plain, and
Alluvial Plain. FP III B (Adjacency Information) clas-
sifies alluvial Plain and flood plain. To identify Bars
connected component-labeling algorithm [11] is used.
Adjacency Information is used to categorize Plains into
Flood Plain or Alluvial Plain. If a plain is near (in
Euclidean sense) to active channels, it is considered to
be ’Flood Plain’, else if it does not have a common
boundary with active channel, it is considered to be
’Alluvial Plain’.
– Bars possess a unique characteristic property of
being enclosed by active channel on all sides.
A connected component-labeling algorithm is em-
ployed to determine if all set of connected pixels
whichever is classified as Plains (detected in FP II
A) are enclosed by an active channel.
– Among different Plains detected (in FP II A and
FP II B) whichever is close to active channel are
classified as flood plains and others are classified
as alluvial plain.
III. R
ESULTS AND DISCUSSION
Training and testing samples from the landform data were
acquired with the help of hand-labeled data to improve the
results of sub-classification stages. Results will be shown
using the same and compared with prior work done by [12],
[13]. The two stages of the landform classification scheme,
as proposed in Section II, have been implemented. The first
stage of super-group characterization consists of a fusion of
three SVM classifiers. The second stage is a hierarchical
organization, where the leaf nodes of the tree indicate the
output for a particular landform, while the intermediate
nodes consist of tailor-made processes, including SVM,
connected component labeling, shape feature detection and
labeling, etc.
Figs. 2, 3 & 4 shows the results obtained for landform
images of different categories: Desertic, Coastal and Fluvial,
where within each figure (a) is the Input image, (b) the
corresponding hand-labeled data, (c) the output obtained
by the unsupervised method [12], (d) the output obtained
from the method proposed in [13] and (e) Output obtained
from our proposed method. Each row in Figs. 2-4 use a
unique color code (label) for each segment detected in the
output map. Unsupervised method works well in case of
landforms covering larger areas and are distinct like salt
flats, Rocky exposure, sea etc. It fails in cases of dunes,
ox-bow and Plains, where our proposed method gives better
classification results. Almost all the landforms have been
correctly identified by our algorithm, since different features
were detected with suitable processing modules and suitable
classifiers were used for identifying the different landform
signatures. For example, template matching algorithm was
used for identifying parabolic dunes and longitudinal dunes
and connected component labeling algorithm was used for
identifying creek and oxbow.
Table III shows the overall accuracy of classification for
each landform obtained at sub-group level. In Table III,
the classification accuracy has been shown for only those
classifiers, for which we were able to compute this measure
using a set of test samples. In other cases this was not
estimated, as we had used trivial image processing methods
(Template Matching, Connected Component Labeling, Adja-
cency Information, etc.) and not classifiers for segmentation
and labeling of the pixels. The overall accuracy of the pro-
posed method (quantitatively measured for a few landforms
and visually compared in other cases using results on raster
images) is better due to the hierarchical organization of a
set of classifiers. Each classifier in our proposed framework,
solves a specific part of the overall problem of classification,
for a small set (2-4) of classes within a limited domain.
Hence the performance of each classifier is quite high
compared to the case of using a single classifier to solve
the complete problem. A large number of classes having
confusing or overlapping feature properties between classes
and large dimension of the feature space would have made
that single classifier provide unsatisfactory performance.
Table IV shows the classification timing taken by different
algorithms for different image sizes. Our proposed method
takes lesser time than the unsupervised method [12], notably
for larger size images. Though our method takes more
time than the method proposed by Gagrani et al. [13], the
performance of our method is superior (see Fig. 2-4).
IV. C
ONCLUSION AND FUTURE WORK
Hand-labeled data was obtained from GIS experts to com-
pare the output produced by our algorithm. The proposed
framework has been developed and tested successfully on all
the samples given to us by GIS experts and the implementa-
tion (Matlab and Visual C++ environments) works up to an
image size of 3000 x 3000. Fusion of DEM (high resolution)
will definitely produce better results, but we intended to
explore this complex problem in the absence of DEM.
This approach gains significance because very accurate and
high-resolution DEMs of unknown areas (especially new
310

Frequently asked questions

Q1. What have the authors contributed in "A hierarchical multi-classifier framework for landform segmentation using multi-spectral satellite images - a case study over the indian subcontinent" ?

The authors propose a hierarchical method for landform classification for identifying a wide variety of landforms occurring over parts of the Indian subcontinent.