You are on page 1of 28

Water Resour Manage (2011) 25:3219–3246

DOI 10.1007/s11269-011-9853-7

Assessing Flood Induced Land-Cover Changes


Using Remote Sensing and Fuzzy Approach
in Eastern Gujarat (India)

Chandra S. Sharma · Mukund D. Behera ·


Atmaram Mishra · Sudhindra N. Panda

Received: 23 November 2009 / Accepted: 22 May 2011 /


Published online: 11 June 2011
© Springer Science+Business Media B.V. 2011

Abstract Natural hazards such as flooding can cause changes in land-cover. The
present study deals with the changes in land-cover in three worst affected districts
(Anand, Vadodara and Kheda) of Gujarat state in India due to severe flood during
2005. The Indian Remote Sensing (IRS) P6 Linear Imaging Self Scanning (LISS) III
satellite imageries of pre- and post-flooding periods were used as sources of informa-
tion for the study area. Three classification approaches (unsupervised ISODATA,
supervised Maximum Likelihood Classifier, and fuzzy rule based) were used to
extract flood induced land-cover information. Results obtained from the above
classification approaches were compared. Soft computing technique such as fuzzy
based image classification gave better separability amongst classes as compared
to hard classification techniques. The accuracy assessment showed that the fuzzy
approach can predict land-cover more accurately than traditional approach and
also showed great potential for dealing with mapping of flood induced land-cover.
Unsupervised classification results for the period October 2004 to October 2005

C. S. Sharma
Agricultural and Food Engineering Department, Indian Institute of Technology,
Kharagpur 721302, West Bengal, India

M. D. Behera
Centre for Oceans, Rivers, Atmosphere and Land Sciences, Indian Institute of Technology,
Kharagpur 721302, West Bengal, India

A. Mishra (B)
Directorate of Water Management (Indian Council of Agricultural Research),
Chandrasekharpur, Bhubaneswar 751023, Orissa, India
e-mail: atmaramm@yahoo.com

S. N. Panda
School of Water Resources, Indian Institute of Technology, Kharagpur 721302,
West Bengal, India
3220 C.S. Sharma et al.

revealed decrease in inland water bodies (14.49%) and agricultural area (6.42%)
while increase in remaining land-cover. During February 2005 to February 2006, all
land-cover classes decreased except agricultural fallow and sparse vegetation. In case
of supervised classification, decreasing trend was observed only in case of agricultural
area (6.78%) during October 2004 to October 2005. Similarly, during February 2005
to February 2006, increase in coastal water bodies (0.73%) and sparse vegetation
(1.7%) was observed where as decreasing trend was noticed in the remaining land-
cover classes. In fuzzy based classification, only decrease in agricultural area (7.09%)
was observed from October 2004 to October 2005, whereas during February 2005 to
February 2006, decrease in area was exhibited in all land-cover classes except coastal
water bodies and sparse vegetation. Change detection indicated interchange of areas
between inland and coastal water bodies and decrease in agricultural area leading to
increase in area of agricultural fallow and sparse vegetation.

Keywords Change detection · Flood · Fuzzy classification · India · Remote sensing ·


Supervised classification · Unsupervised classification

1 Introduction

Water is the finite and most important renewable natural resource required for
agriculture, industries, and domestic purposes. Managing water resources is a major
challenge throughout the world. Space technology plays a crucial role in studies
related to flood, conservation and utilization of water resources, monitoring and
management of water resource dynamics. Improved spatial, spectral and temporal
resolution data from remote sensing found to be appropriate for flood mapping (Jain
et al. 2005), flood monitoring, flood damage assessment (Dewan et al. 2007), land-use
and land-cover change, flood forecasting, validating numerical inundation models
and rainfall run-off analysis (Liu and Smedt 2005; Machado and Ahmad 2007). It can
be used effectively in the areas of flood management, irrigation, assessment of wa-
terlogging (Choubey 1998), snowmelt-runoff forecasts, river configuration, reservoir
sedimentation, vulnerable areas of bank erosion, watershed characteristics and treat-
ment, and drought monitoring. The application of satellite data for flood mapping
began with Landsat-1 (Hallberg et al. 1973; Morrison and White 1976).
Flood is the most hazardous, frequent and wide spread natural disastrous event
throughout the world. More than one-third of the world’s land area is flood prone
affecting about 82% of the world’s population (Dilley et al. 2005). According to
UNDP (2004), about 196 million people in more than 90 countries are exposed to
catastrophic flooding, and between 1980 and 2000, about 170,000 deaths were asso-
ciated with floods worldwide. Destructive floods are common place in lower latitude
regions, particularly in Asia (Kundzewicz et al. 2009).
India being an agricultural country, the prosperity of its agriculture largely de-
pends on rainfall which has significant temporal and spatial variations, unequal sea-
sonal and geographical distribution along with frequent departures from the normal.
Almost 75% of the annual average rainfall occurs during the monsoon season of
four months period (June to September). During monsoon season, about 50% of the
annual average rainfall occurs from few intense storms, which is one of the main
Assessing Flood Induced Land-Cover Changes Using Remote Sensing... 3221

causes of floods in India. Flood prone area in India is increasing dramatically.


Based on the data of 1953 to 2000, Central Water Commission, Ministry of Water
Resources, Government of India reported that the annual average area affected by
flood is 7.563 M ha with a variability ranging from 1.46 M ha in 1965 to 17.5 M ha in
1978 (Mohapatra and Singh 2003).
In India, severe floods occur almost every year in one part of the country or the
other. During last week of June 2005, formation of cyclonic circulation over North
Bay of Bengal led to depression and ended with excessive down pour in central
and southern part of the country. Owing to unprecedented heavy rains, the state
of Gujarat has been experiencing floods and inundation of low lying areas. The
torrential rain that lasted for more than 100 h in the last week of June 2005 caused
severe flood in major rivers of Gujarat state, India. Due to this flood, crops worth
millions of US dollar were destroyed. The death toll was about 123 people in the
state and more than 250,000 people were evacuated. The loss due to flooding was
estimated to be over 80 billion Indian Rupees (about 1.77 billion US dollar). Out
of 25 districts in the state of Gujarat, flood in 2005 affected 19 districts (four slightly
affected, twelve moderately affected and three severely affected). The worst affected
districts were Anand, Vadodara and Kheda. Out of 9,000 affected villages, severely
affected districts (Anand, Vadodara and Kheda) included about 2000 villages. The
cities experienced severe inundation conditions and the rivers were flowing above
danger levels. Due to topography and poor drainage; the villages, roads and other
infrastructures were flooded with excess rain water. Damages due to flood have been
increasing each year resulting in loss of lives, property and agricultural production
as well as affecting normal activities of the area. The socio-economic losses and
environmental changes such as changes in land-cover following extreme flooding
events need to be assessed for future land-use planning, flood mitigating measures
and water resource management. The analysis of spatial extent and temporal pattern
of flood-inundated areas is of prime importance for mitigation of floods. Both
structural and non-structural measures can be undertaken to tackle the problem.
The non-structural measures for mitigation of flood hazards are cost effective as
compared to structural ones (Jain et al. 2006).
For appropriate flood control measures and water resource management, it is
necessary to acquire timely and reliable information about water resources in the
flooded areas, watershed characteristics, river configurations and its behaviour etc.
prior to, during and after the flood. The satellite data are useful in delineating
the boundaries of flood prone zones, identification of the sites ideal for taking up
structural measures to control floods and water resource management, mapping and
assessing the spatial and temporal dynamics of land-use and land-cover (Reger et al.
2007; Serra et al. 2008), and detecting land-use and land-cover changes in the flooded
areas. It has become a critical and universal tool for evaluating the physical attributes
of land and water resources and provides unique opportunity towards comprehensive
monitoring and management of water resources dynamics.
There are many methods available for mapping and monitoring of floods. Numer-
ous researchers have used different methods to successfully map features with dis-
tinct spatial boundaries (Barlow et al. 2006). Several established methods/techniques
are available for classifying images of flood induced remotely sensed data and to
improve the accuracy of land-cover changes such as visual interpretation of satellite
image (Jensen 2005), multi-spectral image classification, band ratioing, contextual
3222 C.S. Sharma et al.

multi-temporal classification and object based classification (Cleve et al. 2008).


Of these, multi-spectral image classification is most commonly used. Different
algorithms and strategies have been developed and tested for various feature ex-
tractions viz., supervised, unsupervised, and hybrid classifications; parametric and
nonparametric classifiers; segmentation; artificial neural networks; fuzzy sets image
segmentation and knowledge-based systems (Wang 1990). Others have reported
object-based classification to map vegetation species or structure in high spatial-
resolution imagery with more detail than conventional pixel-based methods which
has been particularly successful in urban environments (Yu et al. 2006). Herold et al.
(2003) used object-based methods and argued that spatial resolutions better than
5 m are required for urban land-use mapping. Pixel-based classification methods
may be adequate for mapping land-cover over large spatial scales. Of the aforemen-
tioned available methods/techniques, unsupervised [Iterative Self Organised DATA
(ISODATA) Classifier], supervised (Maximum Likelihood Classifier), and fuzzy
rule based classification approaches were used in the present study to extract flood
induced land-cover information of the study area.
Of the various techniques available for change detection, the pre- and post-
classification comparisons have been extensively used (Singh 1989; Coppin et al.
2004). In the present study, the post-classification comparison examines the changes
over time between independently classified land cover data. One of the disadvan-
tages associated with this approach is that the accuracy of the resultant land-cover
change maps depends on the accuracy of the individual classification, meaning that
such techniques are subject to error propagation (Yuan et al. 2005). Despite the
difficulties associated with post-classification comparisons, this technique is most
widely used for identifying land-cover changes (Jensen 2005).
The objective of this present study is to map the flood induced land-cover of the
three worst affected districts (Anand, Vadodara and Kheda) of Gujarat state, India
due to flood in 2005 using different classification approaches and identify the changes
that have taken place using IRS P6 LISS III satellite data. Distribution pattern
of land-cover for pre- and post-flood period was analyzed, both monsoon and dry
seasons. Finally, the changes in land-cover between pre- and post-flood periods were
evaluated using post classification change matrix method.

2 Study Area and Data Used

The study area lies in the Gujarat state located in the western coast of India between
21.1◦ to 24.7◦ N latitude and 68.4◦ to 74.4◦ E longitudes (Fig. 1). The state is divided
into 25 districts and has a coastline of 1,600 km, which forms the western and south
western boundaries. The state can be categorized into two natural regions, i.e., (1)
Gujarat plains; and (2) Kutch–Saurastra peninsula with three typical climate types,
viz., arid, semi-arid, and sub-humid. The study area comprises of three worst affected
districts of Gujarat during flood 2005 and their geographical areas are, Anand
(2,940.8 km2 ), Vadodara (7,549.5 km2 ), and Kheda (4,218.8 km2 ). The mean annual
rainfall of the study area is 863 mm. Temperature of the study area ranges between
27◦ C and 42◦ C. Mahi is the major river that flows through the study area. The area is
flood prone due to its typical topography and poor drainage characteristics. The area
had experienced a number of moderate and severe floods in the past.
Assessing Flood Induced Land-Cover Changes Using Remote Sensing... 3223

INDIA

STUDY AREA

Fig. 1 Location map of the study area

IRS P6 LISS III satellite images with resolution of 23.5 m of paths (93–94) and
rows (56–57) for October–November 2004 (pre-flood monsoon season), February-
March 2005 (pre-flood dry season), October 2005 (Post-flood monsoon season), and
February 2006 (post-flood dry season) were used for the present study. LISS III
has four spectral bands i.e. Green (0.52–0.59 μm), Red (0.62–0.68 μm), NIR (Near
Infrared; 0.77–0.86 μm) and SWIR (Shortwave Infrared; 1.55–1.70 μm).

3 Methodology

In this study, four sets of data (two time periods and two seasons) were used i.e. pre-
flood monsoon and dry seasons; Post-flood monsoon and dry seasons. After radio-
metric correction, rectification, extraction of study area and Normalized Difference
Vegetation Index (NDVI) based separation of vegetated/non vegetated area, three
classification approaches (unsupervised ISODATA, supervised Maximum Likeli-
hood Classifier, and fuzzy rule based) were executed to extract flood induced land-
cover information. Subsequent to classification, accuracy assessment was performed.
This led to area estimation, analysis and comparison of results obtained from
different classification approaches. Finally, post-classification change detection was
carried out to investigate the changes in land-cover between pre- and post-flood
periods using change matrix method. The flowchart showing the methodology used
for this study is given in Fig. 2.

3.1 Radiometric and Geometric Corrections

The study area falls under two different paths. Images acquired for extraction of
the study area had different radiometric properties for both the paths. Path-wise
histogram matching was performed to rectify the images. Thereafter, a master scene
was geometrically corrected with the base map (in the Erdas IMAGINE 9.2 image
processing software) having Root Mean Square (RMS) error below 0.5. Further,
all other scenes were geometrically corrected following image to image registration
method with RMS error below 0.6. Due to RMS error value of <1, geometric
3224 C.S. Sharma et al.

Satellite Data: Two Time Periods (T),


Two Seasons(S)
(4-Sets)(T1S1,T1S2,T2S1,T2S2)

Radiometric Correction (Path-Wise Histogram Matching)


for Different Time Periods

Rectification Using Base Layers

Mosaicking of FCCs; Storage of GCPs

Preparation of Classification Scheme

NDVI-Based Separation of Vegetated / Non-Vegetated Areas

T1S1 T1S2 T2S1 T2S2

Assignment of Initial Clusters Training Set Assignment Fuzzy Membership Assignment

Unsupervised Classification Supervised Classification Fuzzy Rule Based Classification

Accuracy Assessment

Area Estimation of Land-Cover

Comparison between Classification Techniques

Change Analysis of Land-Cover

T1S1 = October–November 2004 (pre-flood monsoon season)


T1S2 = October 2005 (Post-flood monsoon season)
T2S1 = February-March 2005 (pre-flood dry season)
T2S2 = February 2006 (post-flood dry season)
FCC = False Colour Composite
GCP = Ground Control Point
NDVI = Normalized Difference Vegetation Index

Fig. 2 Flow chart showing the methodology

correction was acceptable for the study of land-cover classification and change
detection. RMS error is reported in pixels. It is preferred to tolerate certain amount
of error rather than to take a more complex transformation. As in our case, the
RMS error tolerance was below 0.6; the retransformed pixel was up to 0.6 pixels
away from the source pixel and considered accurate (Leica Geosystems 2005).
Transformation of second order polynomial was performed and resampling was
done using nearest neighbor interpolation method with UTM (Universal Transverse
Mercator) projection system, spheroid WGS 84 (World Geodetic System) and Zone
43. Subsequently mosaicking was performed and the study area was extracted.
Assessing Flood Induced Land-Cover Changes Using Remote Sensing... 3225

3.2 Classification System for Land-Cover

Classification system provides framework for organizing and categorizing informa-


tion that can be extracted from image data. A proper classification scheme includes
classes that are both important for the study and discernible from the data on hand
(Anderson et al. 1976). The level of detailed interpretation from an image primarily
depends on the scale. In this case, level 1 classification system was adopted, which
considers broad classes. Mainly six classes were considered for land-cover clas-
sification i.e. water bodies (inland), water bodies (coastal), agriculture, agricultural
fallow, sparse vegetation (non agricultural and forest), and barren/waste land. Rivers,
streams, reservoir, tanks, and canals were considered under water bodies (inland). As
the south-west part of the study area is along the coast, water bodies (coastal) was
considered to be an important class. The agricultural and horticultural lands were
grouped under the agriculture class. Agricultural fallow consisted of agricultural
lands without any vegetation as well as land just after harvesting of crop. Shrub lands,
barren, and sandy area with shrub canopy greater than 60%, shrubs along the river,
open forest having canopy between 10% and 40% and some moderate forests were
considered under sparse vegetation class. Barren/waste land class consisted of sandy
area along the coast of rivers, dry salt flats, beaches, bare exposed rocks, land without
any vegetation, and settlement areas.

3.3 Normalized Difference Vegetation Index (NDVI) Approach

Wide ranges of enhancement are possible in remotely sensed data (Nagler et al.
2009). These enhancement techniques do not create new information, rather they
highlight information present in the original spectral data. Numerous vegetation
indices have been developed to estimate vegetation cover from the remotely sensed
imagery. Vegetation index is a number that is generated by some combination of
remote sensing bands. The most common spectral index used to evaluate vegetation
cover is the NDVI (Tucker 1979). NDVI-based separation of pixels were done to dis-
tinguish the vegetative and non-vegetative areas. NDVI uses a set of transformation
using Near Infrared (NIR) and Red bands (RED), which is as follows:

N I R − RED
NDVI = (1)
N I R + RED

Ideally values obtained greater than zero represents vegetated area and values less
than or equal to zero represents non-vegetated area. Deletion of negative values
effectively eliminates the non-vegetation information and retains the vegetation
information. During separation it was observed that some other classes are also
going with vegetation class, so a threshold point of 0.15 was taken for separation
of vegetation and non-vegetation classes. This helped in minimizing the error for
getting the training set for classification.

3.4 Classification

Traditional methods of classification are supervised and unsupervised classification


that generate one-pixel-to-one-class mapping. Results of traditional classification
3226 C.S. Sharma et al.

methods assign crisp boundaries to classes. Crisp class assignment is often in-
appropriate in geographical and remote sensing sciences. This is because of the
uncertainties present in the geographical objects. The study area contains a complex
combination of different land-cover classes which often causes mixed pixels in the
remotely sensed data. To quantify uncertainties, it is suggested to use fuzzy approach
based on fuzzy sets (Zadeh 1965).

3.4.1 Unsupervised Classif ication


Unsupervised classification procedure is more automated than supervised clas-
sification (Jensen 2005) and is useful when time is less and high accuracy is not
needed. Unsupervised classification algorithms compare pixel spectral signatures to
the signatures of computer determined clusters and assign each pixel to one of these
clusters (Leica Geosystems 2005; Lillesand et al. 2008). Initialization of the iterative
optimization procedure requires specification of the number of clusters expected.
In practice, the actual or optimum number of clusters to be chosen is not known.
Therefore, it is often chosen conservatively high, having in mind that resulting insep-
arable clusters can be consolidated after the process is completed, or at intervening
iterations, if a merging operation is available. The choice of the initial locations of
the cluster centres is not critical although evidently it will have an influence on the
time it takes to reach a final, acceptable clustering. Since no guidance is available
in general, the logical procedure as mentioned below is followed. The user specifies
a convergence threshold, number of desired classes, and the number of iterations.
Iterations ceases when either the convergence threshold or the maximum number
of iterations is reached. The method is not completely automated, as it requires
the analyst to manually label the resultant spectral classes to desired information
classes (Phillips 1973).
Unsupervised classification was done using ISODATA algorithm in Erdas
IMAGINE 9.2. ISODATA algorithm is a standard unsupervised classifier that uses
minimum spectral distance to assign a cluster for each candidate pixel through a
number of iterations. In the present study, initially 35 clusters were specified followed
by 50 clusters to see whether the ISODATA algorithm has partitioned the feature
space effectively. Requesting more clusters and allowing more iteration would
partition the feature space even better (Jensen 2005), so finally 75 clusters were
specified. Standard deviation was varied from 0.5 to 3, iteration from 5 to 15 as well as
convergence threshold of 0.95 to get desired separability between the classes. These
75 clusters were carefully judged and spectrally similar classes of identical land-
cover types were merged. These merged clusters were evaluated to see whether they
belonged to the desired land-cover classes. Finally, a labelling process was carried out
to generate a thematic land-cover map.

3.4.2 Supervised Classif ication


Supervised classification involves three steps i.e. the training stage, the classification
stage, and the output stage. It requires knowledge of the study area to define
training classes. Regions containing material of interest within a scene are delineated
graphically for use in the supervised classification algorithm. In training stage, the
multi-spectral parameters are extracted for various classes from the training sites
identified in the image. This training set data is used to calculate various statistical
Assessing Flood Induced Land-Cover Changes Using Remote Sensing... 3227

parameters such as mean, standard deviation along with variance and covariance
matrix for carrying out classification. The most common supervised classification
techniques are the Maximum Likelihood Classifier (MLC) for parametric input data
and Parallelepiped Classifier for non-parametric data. The MLC assumes that a pixel
has certain probability of belonging to a particular class. This classifier quantitatively
evaluates both the variance and the covariance of the category spectral response
patterns when classifying an unknown pixel (Shalaby and Tateishi 2007). Since this
classification takes into account the shape of the distribution it is expected to provide
a better accuracy.
Initial step for supervised classification was for training site selection and extrac-
tion of statistics. Homogeneous sample pixels were identified as training pixels (train-
ing set data for 6 classes i.e. inland water bodies, coastal water bodies, agriculture,
agricultural fallow, sparse vegetation, and barren/wasteland) in the image that can
be used as representative samples for each land-cover category to train the algorithm
for locating similar pixels in the image. For each land-cover type, fifteen to twenty
areas of interest were prepared as the signatures of training samples. The reference
data were used to prepare the training signatures. The training areas were created
in order to discriminate the individual classes. Only those training data was selected,
which gave unimodal histograms per class. Multimodal training data was discarded as
it suggests that there are at least two different type of land-cover within the training
area. After obtaining satisfactory discrimination between the classes during spectral
signature evaluation, supervised classification was carried out for all the time period
images of the study area.

3.4.3 Image Classif ication Based on Fuzzy Logic


While classifying an image, generally two kinds of problems are faced. First, in most
of the cases there is no fixed boundary between two land cover classes. Second,
there may be chances of a single pixel containing more than one class. These
problems have lead to the concept of soft classification techniques such as sub-pixel
classification, fuzzy classification (Wang 1990), and image segmentation using fuzzy
c-mean clustering algorithm. In fuzzy classification, fuzzy classifier assigns one pixel
to many classes in varying proportions (Foody and Atkinson 2002). Here, each pixel
can belong to several different classes as it does not have definite boundaries (Jensen
2005).
To handle the concept of “partial truth”, Zadeh (1965) proposed a new theory
called “Fuzzy Sets”. Hard classification procedure may not interpret the boundaries
in an appropriate manner, where as the fuzzy approach, in general, deals with
the vagueness in the boundaries between classes. Fuzzy set theory provides useful
concepts and methods to deal with uncertain information. The set is associated with
a membership function and each element in this set has its own membership value
towards that particular set. The membership values range between 0 and 1. If the
membership value of an element is 0, it means that, it does not belong to that set and
if it is 1, then it belongs to that set completely. But, in crisp sets, the membership value
is either 1 or 0. A fuzzy classification is used to find out uncertainty in the boundary
between classes and to extract the mixed pixel information. This is achieved by
applying a function called “membership function” on remotely sensed images. For
crisp classification, if a pixel P belongs to a class C, then membership function MF
[P, C] = 1, else MF [P, C] = 0. When classes have no definite boundaries, then the
3228 C.S. Sharma et al.

assignment of the pixel to a class is uncertain, which is expressed by fuzzy class


membership function. It takes the value between 0 and 1, such that CLASS (P) =
{C/M [P, C] > 0}. In hard classification, the assignment implies full membership
to single class and no membership to other classes. It is likely that pixel under
investigation has different classes also. Such information is completely lost when the
pixel is assigned to a single class using hard classification. The sum of membership
function values for all classes in each pixel must equal to 1.0. When working with
real remote sensor data, the actual fuzzy partition of spectral space is a family of
fuzzy sets, F1 , F2 ,. . . .., Fm on the universe X such that for every x which is an element
of X (Wang 1990):

0 ≤ fF ≤ 1 (2)


n
f Fi (x) > 0 (3)
i


m
f Fi (x) = 1 (4)
i

Where F1 , F2 , . . . . . . , Fm represents the spectral classes, X represents all pixels in


the data sets, m is the number of classes trained upon, n is the number of pixels, x
is a pixel measurement vector, and f F is the membership function of the fuzzy sets
Fi (1 ≤ i ≤ m). The fuzzy partition may be recorded in the following fuzzy partition
matrix (Wang 1990).
⎛ ⎞
f F1 (x1 ) f F1 (x2 ) . . . . . . f F1 (xn )
⎜ f F2 (x1 ) f F2 (x2 ) . . . . . . f F2 (xn ) ⎟
⎜ ⎟ (5)
⎝ ...... ...... ...... ...... ⎠
f Fm (x1 ) f Fk (x2 ) . . . . . . f Fm (xn )

Where, xi is the ith pixel’s measurement vector (1 ≤ i ≤ n).


Mean and standard deviation values can be taken as parameters for membership
function definition and is also used in the present study. The following two equations
(Eqs. 6 and 7) describe the fuzzy parameters of the training data:

fc (xi ) xi
μ∗c = (6)

n
fc (xi )
i=1


n

T
 fc (xi ) xi − μ∗c xi − μ∗c
∗ i=1
= (7)
c
n
fc (xi )
i=1


∗ the fuzzy mean of training class c is μc ; the fuzzy covariance of training class c
Where,
is c ; the vector value of pixel i is xi ; the membership of pixel xi for training class c is
fc (xi ); T is the transpose of the matrix; n is the total number of pixels of the training
data.
Assessing Flood Induced Land-Cover Changes Using Remote Sensing... 3229

In order to determine the fuzzy mean (Eq. 6) and fuzzy covariance (Eq. 7) of
every training class, the membership of pixel xi needs to be known. The membership
function is defined based on maximum likelihood classification algorithm with fuzzy
mean and fuzzy covariance.

Pc∗ (xi )
fc (xi ) = (8)

m
Pc∗ (xi )
j=1

 −1/2 ∗ −1

e−1/2(xi −μc ) c (xi −μc )
∗ T ∗
Where, Pc∗ (xi ) = (2π )−N/2 c (9)

Where, maximum likelihood probability of pixel xi for training class c is Pc∗ (xi ), the
number of classes is m and the number of the bands is N.
It is desirable that knowledge automation be incorporated into existing fuzzy
systems in order to make the benefits of fuzzy logic available to image classification.
Fuzzy based image classification was executed using Fuzzy Logic Toolbox in Matlab
7.1, which is a compilation of functions built on the Matlab numeric computing
environment and provides tools for creating and editing fuzzy inference systems
within the framework of Matlab. Sugeno method of fuzzy inference system was used.
The Sugeno output membership functions are either linear or constant. A typical rule
in a Sugeno fuzzy model has the form,

If Input 1 = x and Input 2 = y, then Output z = ax + b y + c

For a zero-order Sugeno model, the output level z is a constant (a = b = 0).

3.4.4 Membership Functions


Membership function is the mathematical function, which defines the degree of an
element’s membership in a fuzzy set. Univariate statistics such as mean and standard
deviation were obtained from training data sets used for supervised classification.
These are used as parameters for the fuzzy membership function assignment under
fuzzy rule based classification. Input parameters to define membership function for
individual land-cover class were taken from four individual bands of different time
period images. Gaussian Membership functions were defined in each of four bands
(Green, Red, NIR, and SWIR) for six land-cover classes. Layers of four bands
of satellite data were taken as input for classification. After defining membership
functions, rules were written for classification. Outputs of each rule are combined
into a single fuzzy set by aggregation for decision making. Layers of each band were
extracted from the images, which were used as an input to fuzzy inference system for
classification. Codes for each step of classification process were written in Matlab to
obtain the final land-cover map.

3.5 Accuracy Assessment

Accuracy assessment is an essential and most crucial part of studying image clas-
sification. Land-cover maps derived from remote sensing always contain some sort
3230 C.S. Sharma et al.

of errors due to several factors, which range from classification technique to method
of satellite data capture. In order to wisely use the derived land-cover maps and
the accompanying resource statistics, the errors must be quantitatively explained in
terms of classification accuracy. Whether the output meets expected accuracy or not
is usually determined by the users themselves depending on the type of application
the map product will be used latter. Accuracy levels that are acceptable for certain
application may be unacceptable for others.
The most common and typical method used by researchers to assess classification
accuracy is the use of an error matrix (Congalton and Green 1999; Foody 2002; Fan
et al. 2007). It is also sometimes referred as confusion matrix or contingency table.
These tables produce many statistical measures of thematic accuracy including over-
all classification accuracy, percentage of omission and commission error (Congalton
1991) and the kappa coefficient, an index that estimates the influence of chance
(Congalton and Green 1999; Yuan et al. 2005). Overall accuracy is the simplest and
one of the most popular accuracy measures. User’s accuracy (errors of commission)
represents the probability that a given pixel will appear on the ground as it is
classed, while producer’s accuracy (errors of omission) represents the percentage of
a given class that is correctly identified on the map (Congalton 1991). Kappa is the
proportion of agreement after chance agreement is removed. These values are based
on a sample of error checking pixels of known land-cover that are compared to classi-
fications on the map.
Accuracy assessment of the classified land-cover maps in the present study was
based on reference data, Survey of India topographic maps, and visual inspection of
high resolution images available on the web (Google Earth™). Accuracy assessments
of land-cover maps were carried out (using Erdas Imagine 9.2) by taking 90 randomly
selected points and the results were recorded in an error (confusion) matrix. New
points were selected for each land-cover map before creating the error matrix to
ensure the credibility of accuracy assessment. This was done so that the comparison
between error matrices from different classification techniques could be justified.
The user’s, producer’s and overall accuracies along with the Kappa statistics were
obtained in the accuracy assessment report.

3.6 Change Detection

Remote-sensing techniques have provided a tool through which changes can be


spatially detected, measured, analyzed, and quantified. Many methods of change
detection have been developed to detect land-cover change (Mas 1999; Singh 1989).
In spite of the numerous evaluations of these techniques, no standard techniques
have yet been adopted for all cases (Singh 1989; Macleod and Congalton 1998).
There are many methods of change detection such as image differencing, image ra-
tioing and spectrally based methods which includes tasseled cap analysis, vegetation
index differencing, change vector analysis, and post-classification change analysis or
thematic change analysis. Traditional methods of change detection can be broadly
divided in two categories: pre-classification and post-classification methods (Singh
1989). But post-classification comparison of changes is the most commonly used
method for quantitative analysis. In case of post classification change detection, two
multi-temporal images are classified separately and labelled with proper attributes.
Assessing Flood Induced Land-Cover Changes Using Remote Sensing... 3231

The area of change is then extracted through the direct comparison after obtaining
the classification results (Howarth and Wickware 1981). Owojori and Xie (2005)
have demonstrated the potential of post classification for accurate land-cover change
assessment. In this study, the post-classification change matrix method was used for
change detection in land-cover, as change matrix gives to–from change information
about the flood induced land-cover for two different time periods.

4 Results and Discussion

Each classification approach produced respective thematic land-cover maps. The


selection of classification approaches has impact on quantitative spatial extent of
land-cover (Table 1) as discussed below:

4.1 Unsupervised Classification

Area under different land-cover for all the time periods was computed (Table 1).
The estimated area as the percentage of total study area under this classification
approach for inland water bodies ranged from 0.86% to 1.37%; coastal water bodies
from 2.65% to 2.86%; agriculture from 38.68% to 41.80%; agricultural fallow from
7.42% to 9.68%; sparse vegetation from 28.56% to 29.64%; and barren/wasteland
from 17.59% to 17.99% during the period October 2004 to February 2006. It is
observed that area occupied by inland water bodies decreased in both the seasons
while coastal water bodies increased (6.98%) between October–November 2004
and October 2005 and remained almost same between February–March 2005 and
February 2006. Generally area under inland water bodies is expected to increase
in post-flood period. However, results obtained from this classification showed
decrease in area. This might be due to misclassification of pixels for inland water
bodies. One of the main reasons for misclassification is the similar spectral response
of inland and coastal water bodies. Therefore, careful judgment is necessary before
merging and labeling of clusters for each class.
Another significant change was the decline in agricultural area (6.42%) from
October–November 2004 to October 2005, which might be due to the effect of
flood occurred just 3 months before i.e., in June–July 2005. The croplands in the
three affected districts were destroyed due to severe inundation conditions because
of flood. Conversely, agricultural fallow increased (16.01%) between October 2004
and October 2005 and remained approximately same between February 2005 and
February 2006, indicating reduction in the effect of flood in the dry season. A small
increase in sparse vegetation (3.82%) and barren/wasteland (1.74%) was observed
between monsoon of 2004 and 2005. It also indicated an increase of 1.60% for sparse
vegetation and decrease of barren/wasteland by 1.37% between winter of 2005 and
2006. Unsupervised classification gave expansive overview of land-cover of the study
area with some misclassifications.
It was observed that inland water bodies and agricultural classes were inaccurately
classified. Similarly, some of the sparse vegetation class got merged with agriculture
class due to similar spectral response and was not separated properly. Some of the
barren/wasteland class was also incorrectly classified as agricultural fallow land. This
Table 1 Land-cover in the study area
3232

Sl. No. Land-cover type Area (km2 ) Percentage of total area (%)
October–November February–March October February October–November February–March October February
2004 2005 2005 2006 2004 2005 2005 2006
1. Water bodies (inland) 205.33a 153.98 175.57 129.17 1.37 1.02 1.17 0.86
121.64b 114.42 128.70 106.36 0.81 0.76 0.86 0.71
141.82c 140.61 159.19 130.22 0.94 0.93 1.06 0.87
2. Water bodies (coastal) 398.70a 430.42 426.51 429.02 2.65 2.86 2.84 2.85
439.95b 442.77 446.36 446.00 2.93 2.94 2.97 2.96
424.57c 425.58 426.54 430.79 2.82 2.83 2.84 2.86
3. Agriculture 6,213.08a 6,288.76 5,814.06 6,273.37 41.36 41.80 38.68 41.70
6,225.25b 6,280.20 5,802.99 6,260.83 41.42 41.74 38.61 41.62
6,256.50c 6,292.51 5,813.22 6,273.40 41.62 41.83 38.74 41.70
4. Agricultural fallow 1,254.43a 1,116.04 1,455.28 1,124.44 8.35 7.42 9.68 7.47
1,287.72b 1,127.13 1,441.07 1,112.09 8.57 7.49 9.59 7.39
1,307.09c 1,139.69 1,456.69 1,123.63 8.70 7.58 9.71 7.47
5. Sparse vegetation 4,290.64a 4,372.32 4,454.63 4,442.30 28.56 29.06 29.64 29.53
4,272.07b 4,386.15 4,466.16 4,460.84 28.42 29.15 29.71 29.65
4,255.50c 4,372.26 4,454.21 4,443.61 28.31 29.06 29.68 29.54
6. Barren/waste land 2,658.49a 2,682.87 2,704.62 2,646.09 17.70 17.83 17.99 17.59
2,684.04b 2,693.72 2,745.39 2,658.27 17.86 17.91 18.27 17.67
2,645.19c 2,673.74 2,696.92 2,642.74 17.60 17.77 17.97 17.57
a Land-cover obtained from unsupervised classification
b Land-cover obtained from supervised classification
c Land-cover obtained from fuzzy classification
C.S. Sharma et al.
Assessing Flood Induced Land-Cover Changes Using Remote Sensing... 3233

problem occurred due to spectral in-homogeneity of these classes. For example, a


variety of species of bushes and grasses on sparse vegetation, barren/waste land, and
agricultural fallow caused diverse spectral response. Unwanted vegetation (grasses
and shrubs) grew in the flood inundated area after flood water receded. Those
vegetations on agricultural fallow, sparse vegetation and barren/waste lands gave
similar spectral response, which led to misclassification between them.

4.2 Supervised Classification

The ambiguities as mentioned in the above paragraph were taken care of by the
supervised classification and it gave better results as compared to the unsupervised
classification. The final classified maps of the study area for different time periods
are shown in Fig. 3.

October-November 2004 February-March 2005

October 2005 February 2006

Water bodies (Inland) Water bodies (Coastal) Agriculture


Agricultural Fallow Sparse Vegetation Barren / Wasteland

Fig. 3 Land-cover map of the study area for different time periods
3234 C.S. Sharma et al.

The estimated area as the percentage of total area under this classification
approach for inland water bodies ranged from 0.71% to 0.86%; coastal water bodies
from 2.93% to 2.97%; agriculture from 38.61% to 41.74%; agricultural fallow from
7.39% to 9.59%; sparse vegetation from 28.42% to 29.71%; and barren/wasteland
from 17.67% to 18.27% during the period October 2004 to February 2006 (Table 1).
There is an increase in inland water bodies (5.80%) during the period October–
November 2004 to October 2005. It is primarily due to topography and poor drainage
of the study area, which caused the rainwater to change its course and flood the
villages, roads and other infrastructures. Over flowing of Vishwamitri (Vadodara
district), Shedhi, and Vatrak (Kheda district) rivers along with small rivulets flowing
from Vadodara also led to the inundation of the study area. It was observed that after
flood water receded, water was still stagnating in the study area, which indicated
lack of adequate drainage system to dispose the rainwater from intense storm. In
urban areas, particularly in Vadodara, concrete construction (buildings, roads etc.)
especially in low lying area and along the river bed has led to restricted movement
of water into the soil and enhanced overland flow. Conversely, in February 2006 the
area under inland water bodies decreased (7.04%) as compared to February 2005.
Area under coastal water bodies increased in both the time periods. Decrease in
agricultural area (6.78%) was observed in October 2005, just after the occurrence
of flood and is mainly due to flood. The losses included mainly agricultural (pulses,
jowar, soyabean and paddy) and horticultural crops.
In winter (February 2005 and February 2006) approximately same agricultural
area was observed indicating the decrease in the effect of flood. Agricultural fallow
increased by 11.91% (153.35 km2 ) in October 2005. On the contrary, a decrease of
1.33% was noticed in February 2006 as compared to February 2005 indicating that
the lands were put under cultivation. Sparse vegetation area increased by 4.54%
(194.09 km2 ) in monsoon season showing the effect of flood; but in winter the
increase was 1.70% indicating that there was gradual reduction in sparse vegetation
area from just after flood to February 2006.
The area under inland water bodies decreased (17.66–40.76%) while coastal
water bodies (2.87–10.35%) and barren/waste land (0.40–1.51%) increased in all the
time periods in case of supervised classification when compared with unsupervised
classification. Results from supervised classification showed increased agricultural
area (0.2%) and decreased sparse vegetation area (0.43%) in October–November
2004 against unsupervised classification. Conversely, in other time periods decrease
in agricultural (0.14% to 0.20%) and increase in sparse vegetation area (0.26% to
0.42%) was observed. In case of agricultural fallow, increase of area in pre-flood
seasons and decrease in post-flood season was observed in supervised classification
when compared with unsupervised classification. To take account of some fuzziness
in classifying the boundary pixels as well the mixed pixels, fuzzy based classification
was done.
The above obtained results indicate possible scope for flood mitigations and
water management practice measures to reduce the effect of flood in the study
area in future. The measures should consist of an optimal, site specific structural,
non-structural or both. This includes construction and improvement of existing
embankments in low-lying areas, enlarging of reservoirs; improvement in drainage
systems and dam safety; and watershed management practices to reduce runoff and
erosion.
Assessing Flood Induced Land-Cover Changes Using Remote Sensing... 3235

4.3 Fuzzy Classification

Individual bands of input parameters that were used to define membership functions
(Figs. 4 and 5) showed an overlap between sparse vegetation and agricultural fallow,
which was due to some similar characteristics in spectral response of the classes in
the wave length range of 0.50–0.59 μm. Since the classification procedure consid-
ered multivariate image statistics, this overlap was taken care by other bands for
better separability between land-cover classes. Agriculture class was well separated
specially in NIR band where vegetation cover played an important role with some
overlap between agriculture and agricultural fallow. Barren/wasteland with high
brightness value was better separated in all the bands except NIR band. Water has
high reflectance value in green band and minimum in NIR band. It was observed that
separability of inland and coastal water bodies was acceptable in all the bands.

Feb-Mar 2005 Feb 2006


300 Band 1 300 Band 1
Brightness Value
Brightness Value

250 250
200 200
150 150
100 100
50 50
0 0

Agricultural

Vegetation

Wasteland
Agricultural

Vegetation

Wasteland

Water bodies

Water bodies

Agriculture
Water bodies

Water bodies

Agriculture

Barren/
Barren/

(Coastal)

Sparse
(Coastal)

Sparse

(Inland)

Fallow
(Inland)

Fallow

300
Brightness Value

Band 2 300 Band 2


Brightness Value

250 250
200 200
150 150
100 100
50 50
0 0
Agricultural

Vegetation

Wasteland
Water bodies

Water bodies

Agriculture
Agricultural

Vegetation

Wasteland

Barren/
Water bodies

Water bodies

Agriculture

Barren/

(Coastal)

Sparse
(Coastal)

(Inland)
Sparse

Fallow
(Inland)

Fallow

300 Band 3 Band 3


Brightness Value

300
Brightness Value

250 250
200 200
150 150
100 100
50 50
0 0
Agricultural

Vegetation

Wasteland

Agricultural

Vegetation

Wasteland
Water bodies

Water bodies

Agriculture

Water bodies

Water bodies

Agriculture
Barren/

Barren/
(Coastal)

(Coastal)
Sparse

Sparse
(Inland)

(Inland)
Fallow

Fallow

Band 4 300 Band 4


Brightness Value

250
Brightness Value

250
200
200
150
150
100 100
50 50
0 0
Agricultural

Vegetation

Wasteland

Agricultural

Vegetation

Wasteland
Water bodies

Water bodies

Agriculture

Water bodies

Water bodies

Agriculture
Barren/

Barren/
(Coastal)

(Coastal)
Sparse

Sparse
(Inland)

(Inland)
Fallow

Fallow

Fig. 4 Class overlaps in different bands for Feb–Mar 2005 and Feb 2006
3236 C.S. Sharma et al.

Oct-Nov 2004 Oct 2005


Band 1 Band 1
Brightness Value

200

Brightness Value
200
150 150
100 100
50 50
0 0

Vegetation

Vegetation
Water bodies

Water bodies

Agriculture

Agricultural

Wasteland

Water bodies

Water bodies

Agriculture

Agricultural

Wasteland
Barren /

Barren /
Sparse

Sparse
(Coastal)

(Coastal)
(Inland)

(Inland)
Fallow

Fallow
150
Brightness Value

Band 2 Band 2

Brightness Value
200
100 150
100
50
50
0 0
Vegetation
Water bodies

Water bodies

Agriculture

Agricultural

Wasteland

Vegetation
Water bodies

Water bodies

Agriculture

Agricultural

Wasteland
Barren /
Sparse
(Coastal)

Barren /
Sparse
(Coastal)
(Inland)

Fallow

(Inland)

Fallow
200
Brightness Value

Band 3 Band 3
Brightness Value

150 200
150
100
100
50 50
0 0
Vegetation

Vegetation
Water bodies

Water bodies

Agriculture

Agricultural

Wasteland

Water bodies

Water bodies

Agriculture

Agricultural

Wasteland
Barren /
Sparse
(Coastal)

Barren /
Sparse
(Coastal)
(Inland)

Fallow

(Inland)

Fallow
200
Brightness Value

Band 4 200 Band 4


Brightness Value

150
150
100
100
50
50
0
0
Vegetation
Water bodies

Water bodies

Agriculture

Agricultural

Wasteland

Vegetation
Barren /
Sparse

Water bodies

Water bodies

Agriculture

Agricultural

Wasteland
(Coastal)
(Inland)

Fallow

Barren /
Sparse
(Coastal)
(Inland)

Fallow

Fig. 5 Class overlaps for different bands in Oct–Nov 2004 and Oct 2005

The output images obtained from Fuzzy classification process showed improved
separability among land-cover classes as compared to supervised classification. Com-
parison showed (Fig. 6) that in supervised classification, some area of barren land
was getting classified under sparse vegetation class, whereas in fuzzy classification
the separability was quite satisfactory. Similar trend was seen between inland and
coastal water bodies. Separation among agriculture, agricultural fallow and sparse
vegetation was also better. The separability among the land-cover further increased
when the standard deviation was reduced to half of its original value but in most of
the cases there was an overlap amongst classes and some unclassified pixels were also
generated. Study area consisted of 17,914,038 pixels. Percentage and amount of pixels
occupied by different land-cover from supervised and fuzzy classification (Table 2)
was in proportion to area occupied by different land-cover. There was an increase in
amount of classified pixels for inland water bodies (24,051–36,641), agricultural area
(14,658–37,245), and agricultural fallow area (13,741–23,086) in fuzzy classification
Assessing Flood Induced Land-Cover Changes Using Remote Sensing... 3237

a b c
Fig. 6 A zoomed view of land-cover classes to compare the results of techniques followed: a original
FCC, b supervised classification and c fuzzy classification

as compared to supervised for different time periods. On the contrary, decrease


in amount of classified pixels was also observed for coastal water bodies (18,330–
22,812), sparse vegetation (5,788–20,517), and barren/waste land (18,492–52,649).
Mixed pixels were taken care by fuzzy classification as observed from the result.
Area under different land-cover follows similar trend as obtained from supervised
classification. In fuzzy based classification, an increase in area (16.59% to 23.69%) for
inland water bodies was observed in comparison to supervised classification and the
increased area have come mostly from coastal water bodies. Agricultural area also
increased (0.18% to 0.50%) in different seasons, which came primarily from sparse
vegetation. Similar trend was seen for agricultural fallow in which the increased area
has come mostly from agriculture.
The land-cover area statistics derived from the supervised and fuzzy approaches
showed small differences in spatial extent as compared to the unsupervised ap-
proach. Automated clustering technique (unsupervised) often failed or overesti-
mated/underestimated, mainly for water bodies, agricultural fallow and barren/
wastelands. High contrast was not observed among the land-cover spatial statistics
of different classification approaches. In this case, the fuzzy approach has shown sat-
isfactory results. However, all classifiers showed little differences in the spatial extent
(within ±1% of study area) of the land-cover classes.

4.4 Accuracy Assessment

Before using the classification result for change detection, it is important to test the
result. The most common way of accuracy assessment is by creating an error matrix
or confusion matrix (Foody 2002), which has been used in the present study. Chances
of error or misclassification in October 2005 image was highest as lots of unwanted
Table 2 Percentage of classified pixels for different land-cover in supervised and fuzzy classification
3238

Land-cover October–November 2004 February–March 2005 October 2005 February 2006


Supervised Fuzzy Supervised Fuzzy Supervised Fuzzy Supervised Fuzzy
classification classification classification classification classification classification classification classification
Water bodies 0.81 (144974) 0.94 (169026) 0.76 (136245) 0.93 (167431) 0.86 (153389) 1.06 (190030) 0.71 (126648) 0.87 (155059)
(inland)
Water bodies 2.93 (524347) 2.82 (506016) 2.94 (527226) 2.83 (506757) 2.97 (531986) 2.84 (509174) 2.96 (531072) 2.86 (512961)
(coastal)
Agriculture 41.42 (7419454) 41.62 (7456699) 41.74 (7478119) 41.83 (7492777) 38.61 (6916191) 38.74 (6939418) 41.62 (7455054) 41.70 (7470022)
Agricultural 8.57 (1534746) 8.70 (1557832) 7.49 (1342125) 7.58 (1357081) 9.59 (1717514) 9.71 (1738895) 7.39 (1324216) 7.47 (1337957)
fallow
Sparse 28.42 (5091591) 28.31 (5071842) 29.15 (5222788) 29.06 (5206248) 29.71 (5322914) 29.68 (5317126) 29.65 (5311725) 29.54 (5291208)
vegetation
Barren/waste 17.86 (3198926) 17.60 (3152623) 17.91 (3207535) 17.77 (3183744) 18.27 (3272044) 17.97 (3219395) 17.67 (3165323) 17.57 (3146830)
land
Value in parentheses ( ) indicates number of pixels
C.S. Sharma et al.
Assessing Flood Induced Land-Cover Changes Using Remote Sensing... 3239

vegetation grew on agricultural fallow, sparse vegetation, and barren/wasteland area


after the flood water receded. Maximum probability of inaccuracy was possible in
classification of image for this time period. So the accuracy assessment of classified
land-cover map for this time period is presented. Since the results obtained from
accuracy assessment for the other time periods was comparatively similar, they were
omitted.
The error matrix correspondent to the final result from different classification
process is presented in Tables 3, 4 and 5. In this study, the fuzzy approach appeared
to be the best approach. Fuzzy-based classification gave higher overall accuracy
and kappa (K∧ ) coefficient than the other classification techniques. The overall
classification accuracy and kappa statistics (K∧ ) of fuzzy-based classification were
88.82% and 0.87 (Table 3); for supervised classification were 86.76% and 0.84
(Table 4) and for unsupervised classification the respective values were 81.69%
and 0.78 (Table 5), respectively. This meets the requirement that K∧ values >0.80
(i.e., >80%) represent strong agreement, values between 0.40 and 0.80 (i.e., 40% to
80%) represent moderate agreement and values <0.40 (i.e., <40%) represent poor
agreement (Landis and Koch 1977; Jensen 2005).
Fuzzy approach showed an overall accuracy of 88.82%, which is close to the
overall accuracy (86.76%) of the supervised approach. The fuzzy approach dealt
with mixed pixel problems and the heterogeneous representation of land-cover. The
unsupervised approaches produced lower accuracy (81.69%). Due to the complexity
in land-cover, the unsupervised approach formed several class clusters in image,
creating difficulties in interpretation. However, the unsupervised approach provided
better insight to identify certain classes such as agriculture and sparse vegetation in
some time periods. The kappa statistics presented a somewhat clearer picture. The
kappa coefficient shows that the fuzzy based approach can reduce most of the errors
during the classification process. The kappa for the Fuzzy approach was 0.87 (87%
reduction of error), which is a bit better than the kappa for the supervised approach
of 0.84, with a difference of 2.06% in overall classification accuracy between them. A

Table 3 Error matrix for fuzzy classification


Water Water Agriculture Agricultural Sparse Barren/ User’s Producer’s
bodies bodies fallow vegetation waste accuracy accuracy
(inland) (coastal) land
Water 12 0 0 0 0 0 100.00 92.31
bodies
(inland)
Water 1 14 0 0 0 0 93.33 100.00
bodies
(coastal)
Agriculture 0 0 16 1 2 0 84.21 100.00
Agricultural 0 0 0 11 0 1 91.67 73.33
fallow
Sparse 0 0 0 1 15 1 88.24 83.33
vegetation
Barren/waste 0 0 0 2 1 12 80.00 85.71
land
Overall accuracy = 88.82%, Kappa statistics = 0.87
3240 C.S. Sharma et al.

Table 4 Error matrix for supervised classification


Water Water Agriculture Agricultural Sparse Barren/ User’s Producer’s
bodies bodies fallow vegetation waste accuracy accuracy
(inland) (coastal) land
Water 14 2 0 0 0 0 87.50 93.33
bodies
(inland)
Water 1 11 0 0 0 0 91.67 84.62
bodies
(coastal)
Agriculture 0 0 16 2 2 0 80.00 88.89
Agricultural 0 0 0 10 0 1 90.91 76.92
fallow
Sparse 0 0 2 0 14 0 87.50 82.35
vegetation
Barren/waste 0 0 0 1 1 13 86.67 92.86
land
Overall accuracy = 86.76%, Kappa statistics = 0.84

higher difference of kappa value (0.06) and overall accuracy percentage (5.07%) was
observed between the supervised and unsupervised approach. This signifies that the
supervised approach performed better than the unsupervised in mapping the flood
induced land-cover.
The overall accuracy and kappa coefficient value only represent the average
result. It is still difficult to determine which approach showed refined mapping
of land-cover at the class level. Fuzzy and supervised approaches exhibited high
(>80%) producer’s accuracies or low omission errors in all classes except agricultural
fallow. User’s accuracy for fuzzy and supervised classification also exhibited high
value (>80%) for all land-cover classes. Mapping of land-cover improved to a

Table 5 Error matrix for unsupervised classification


Water Water Agriculture Agricultural Sparse Barren/ User’s Producer’s
bodies bodies fallow vegetation waste accuracy accuracy
(inland) (coastal) land
Water 11 2 0 0 0 0 84.62 78.57
bodies
(inland)
Water 3 12 0 0 0 0 80.00 85.71
bodies
(coastal)
Agriculture 0 0 15 2 3 0 75.00 78.95
Agricultural 0 0 2 11 0 0 84.62 78.57
fallow
Sparse 0 0 2 0 13 1 81.25 76.47
vegetation
Barren/waste 0 0 0 1 1 11 84.62 91.67
land
Overall accuracy = 81.69%, Kappa statistics = 0.78
Assessing Flood Induced Land-Cover Changes Using Remote Sensing... 3241

great extent by the fuzzy approach, showing over 84% user’s accuracies. The fuzzy
approach, with an user’s accuracy of 100%, appeared to be better than the supervised
approach (user’s accuracy of 87.50%), an increase of 12.50% for extracting inland
water bodies. The user’s accuracies show that the supervised approach estimates the
barren/wasteland class better than the other approaches. In the supervised approach,
despite a producer’s accuracy of 93.33% for the inland water bodies, there was
actually 87.50% user’s accuracy, which means at least 5.83% of the inland water
bodies were classified wrong.
Aforementioned results obtained from the accuracy assessment indicate that
classified land-cover maps have sufficient accuracy for post-classification change
detection. Perusal of the error matrix reveals that there is diminutive confusion
in discriminating agriculture from agricultural fallow and sparse vegetation; and
barren/wasteland from agricultural fallow and sparse vegetation. Similar diminu-
tive confusion also existed among agricultural fallow and barren/wasteland; sparse
vegetation and agriculture. Keeping in view the heterogeneity of different classes
present in the study area, the classification accuracies obtained is in fair agreement.
Although the classified result has some misinterpreted classes but overall it shows
good result in terms of overall accuracy, user’s accuracies, producer’s accuracy, and
kappa coefficient to perform change detection.

4.5 Change Detection

Change detection analysis was carried out for pre- and post-flood data of same sea-
son, i.e., between October 2004 and October 2005 (monsoon season); and between
February 2005 and February 2006 (dry season). Change maps for both the seasons
are shown in Figs. 7 and 8. Out of three districts, Anand was the most affected

Water bodies (Inland)

Water bodies (Coastal)


Agriculture
Agricultural Fallow

Sparse Vegetation
Barren/Wasteland

Change

Fig. 7 Land-cover map for October 2004–October 2005


3242 C.S. Sharma et al.

Water bodies (Inland)

Water bodies (Coastal)

Agriculture

Agricultural Fallow
Sparse Vegetation

Barren / Wasteland
Change

Fig. 8 Land-cover map for February 2005–February 2006

one due to its topography and location. Flood plains along the river were mostly
affected and exhibited major changes. The change matrix analysis was carried out
using supervised classification results. The land-cover change matrix of the study area
for the monsoon and dry season is shown in Tables 6 and 7, respectively.
From the change matrix (Tables 6 and 7), changes to–from were analyzed. There
was an increase in inland water bodies in October 2005, i.e. 7.09 km2 of coastal water
bodies area got converted to inland water bodies (Table 6). Increase in inland water
bodies was mainly due to the flood inundation caused by poor drainage and change
in water course. Optimal water resource management practice such as improvement

Table 6 Change Detection for October–November 2004 and October 2005


Land-cover October–November 2004 (Area km2 )
Water Water Agriculture Agricultural Sparse Barren/
bodies bodies fallow vegetation waste
(inland) (coastal) land
October 2005 (area km2 )
Water 116.63 7.09 2.89 1.20 0.38 0.51
bodies
(inland)
Water 2.47 431.28 0.57 5.04 3.60 3.40
bodies
(coastal)
Agriculture 0.30 0.13 5627.38 47.95 119.43 7.80
Agricultural 0.49 0.52 212.43 1030.19 164.80 32.64
fallow
Sparse 0.71 0.30 306.80 103.45 3912.21 142.69
vegetation
Barren/waste 1.04 0.63 75.19 99.89 71.65 2496.99
land
Assessing Flood Induced Land-Cover Changes Using Remote Sensing... 3243

Table 7 Change Detection for February -March 2005 and February 2006
Land-cover February–March 2005 (area km2 )
Water Water Agriculture Agricultural Sparse Barren/
bodies bodies fallow vegetation waste
(inland) (coastal) land
February 2006 (area km2 )
Water 95.24 0.91 8.19 0.18 1.17 0.67
bodies
(inland)
Water 9.82 429.7 1.04 1.25 2.24 1.95
bodies
(coastal)
Agriculture 5.79 3.42 5964.8 73.44 200.62 12.76
Agricultural 1.47 4.32 80.37 934.04 83.33 8.56
fallow
Sparse 0.23 3.01 214.57 111.81 3993.3 137.92
vegetation
Barren/waste 1.87 1.41 11.23 6.41 105.49 2531.86
land

in land drainage systems, embankments, and watershed management practices is


required for flood mitigation in the study area. Agricultural area decreased in
October 2005, i.e. 212.43 km2 went to agricultural fallow and 306.80 km2 to sparse
vegetation. Agricultural loss in the study area that resulted from flood can be
minimized with proper management practices discussed above. Increase of area
under sparse vegetation was also observed with area appeared to coming from
agricultural land (306.80 km2 ), agricultural fallow (103.45 km2 ) and barren/wasteland
(142.69 km2 ). Decrease in area under inland water bodies and minimal increase in
coastal water bodies (Table 7) was observed between February 2005 and February
2006. Agriculture and agricultural fallow land decreased but the decrease was not
significant. 80.37 km2 areas became agricultural fallow and on the other hand
73.44 km2 area of agricultural fallow land was used for cultivation. Area covered
under sparse vegetation increased, which came from agriculture (214.57 km2 ), agri-
cultural fallow (111.81 km2 ) and barren wasteland (137.92 km2 ). There has been
interchange of areas between agriculture, agricultural fallow, sparse vegetation and
barren/wasteland in both the time periods. It can be inferred from above discussion
that between February 2005 and February 2006, land-cover change due flood in the
study area was relatively stable as compared to October 2004 and October 2005.
Results obtained from change detection can be considered satisfactory as evident
from the accuracy assessments of land-cover map.
In supervised classification, simple difference in area of individual land-cover
classes observed in both the time periods with respect to the total study area (Table 1)
showed decrease of around 2% in agriculture and increase of approximately 2% in
sparse vegetation between October 2004 and October 2005. Similarly, decrease of
around 1% in agriculture and increase of approximately 1% in sparse vegetation
area was observed between February 2005 and February 2006. But in fuzzy based
classification (Table 1), decrease of around 3% in agriculture and increase of ap-
proximately 1% in agricultural fallow, sparse vegetation, and barren/wasteland area
3244 C.S. Sharma et al.

was observed between October 2004 and October 2005. Minimal changes in land-
cover were observed between February 2005 and February 2006 in the study area.

5 Conclusions

The flood often leaves its imprint on the surface in the form of soil moisture
anomalies, inundated areas, soil scours, stressed vegetation, debris lines, and other
indicators of the flooded area for days, weeks, or even months after the flood water
recedes, which helps in the study of land-cover change. Remote sensing can be a
very effective tool towards flood mapping and monitoring the change. Its integration
with monitoring and mitigation mechanisms is helpful for hazard reduction. In this
case, flood induced land-cover mapping using three different classification approach
(unsupervised ISODATA, supervised Maximum Likelihood Classifier, and fuzzy
rule based) and land-cover change with respect to pre- and post-flooding periods
for monsoon and dry season was studied using IRS P6 LISS III imageries.
The effect of flood was more prominently observed in monsoon season (October
2004 and October 2005) as compared to winter or dry season (February 2005 and
February 2006). Among all the land-cover classes, agriculture was most affected due
to flood. It occupied around 42% of the total study area in the pre-flood monsoon
period which decreased to around 39% in the post-flood monsoon period. Decrease
in agriculture area was around 6.42–7.09%, which has resulted in increase in area for
agriculture fallow, sparse vegetation. The losses were primarily due to poor drainage
and post-flood inadequate water management practices in the study area.
Agriculture was classified more accurately in all the classification approaches
with small variation. Generally area under inland water bodies should increase in
post-flood period (October 2005). Similar trend was observed in results obtained
from supervised (5.8%) and fuzzy based classification (12.25%). On the contrary,
reverse trend was observed in unsupervised classification (14.49%), indicating mis-
classification. So, careful judgment of spectrally similar classes is essential before
merging and labeling of clusters in unsupervised classification. Results revealed that
some classes can be separated even better by using soft computing technique such
as fuzzy logic. The fuzzy approach presented more accurate results and exhibited
better separability of classes as compared to traditional pixel-based classification
approaches (supervised and unsupervised) for improving the land-cover information.
Further, the accuracy assessment showed that the fuzzy approach can predict land-
cover more accurately. It has great potential in mapping of flood induced land-cover
with low difference in the errors of omission and commission.
Present study showed usefulness of remote sensing in mapping of flood induced
land-cover and change detection. It also showed possible scope for flood mitigation
and optimal water management practice measures to be undertaken to reduce the
effect of flood in the study area. Different factors such as agricultural practices,
drainage, unplanned urbanization, soil erosion, sediment accumulation, and river
bed aggradations along with site specific structural measures should be taken into
account for better flood mitigation and water resource management.

Acknowledgements The authors are thankful to the Directorate of Water Management (ICAR),
Bhubaneswar, Orissa (India) for providing necessary remote sensing and other field data. In
addition, the help received from the Spatial Analysis and Modeling Laboratory, CORAL, Indian
Institute of Technology, Kharagpur (India) is highly acknowledged.
Assessing Flood Induced Land-Cover Changes Using Remote Sensing... 3245

References

Anderson JR, Hardy EE, Roach JT, Witmer RE (1976) A land use and land cover classification
system for use with remote sensor data. US Geological Survey Professional Paper No 964,
Washington, DC
Barlow J, Franklin SE, Martin Y (2006) High spatial resolution satellite imagery, DEM derivatives,
and image segmentation for the detection of mass wasting processes. Photogramm Eng Remote
Sensing 72(6):687–692
Choubey VK (1998) Assessment of waterlogging in Sriram Sagar command area, India, by remote
sensing. Water Resour Manag 12(5):343–357
Cleve C, Kelly M, Kearns F R, Moritz M (2008) Classification of the wildland–urban interface:
a comparison of pixel- and object-based classifications using high-resolution aerial photography.
Comput Environ Urban Syst 32(4):317–326
Congalton RG (1991) A review of assessing the accuracy of classifications of remotely sensed data.
Remote Sens Environ 37(1):35-46
Congalton RG, Green K (1999) Assessing the accuracy of remotely sensed data: principles and
practices. Lewis, Boca Raton, p 137
Coppin P, Jonckheere I, Nackaerts K, Muys B, Lambin E (2004) Digital change detection methods
in ecosystem monitoring: a review. Int J Remote Sens 25(9):1565–1596
Dewan AM, Nishigaki M, Kumamoto T (2007) Evaluating flood hazard for land-use planning in
Greater Dhaka of Bangladesh using remote sensing and GIS techniques. Water Resour Manag
21(9):2101–2116
Dilley M, Chen RS, Deichmann U, Lerner-Lam AL, Arnold M with Agwe J, Buys P, Kjekstad O,
Lyon B, Yetman G (2005) Natural disaster hotspots: a global risk analysis. International Bank for
Reconstruction and Development/The World Bank and Columbia University, Washington, DC
Fan F, Weng Q, Wang Y (2007) Land use/land cover change in Guangzhou, China, from 1998 to
2003, based on Landsat TM/ETM+ imagery. Sensors 7(7):1323–1342
Foody GM (2002) Status of land cover classification accuracy assessment. Remote Sens Environ
80(1):185–201
Foody GM, Atkinson PM (2002) Uncertainty in remote sensing and GIS. Wiley, Chichester
Hallberg G R, Hoyer B E, Rango A (1973) Application of ERTS-1 imagery to flood inundation
mapping. In: Symposium on significant results obtained from earth resources satellite 1 1A,
NASA special publication no 327, pp 745–753
Herold M, Liu XH, Clarke KC (2003) Spatial metrics and image texture for mapping urban land-use.
Photogramm Eng Remote Sensing 69(9):991–1001
Howarth P, Wickware G (1981) Procedures for change detection using Landsat digital data. Int
J Remote Sens 2(3):277–291
Jain SK, Singh RD, Jain MK, Lohani AK (2005) Delineation of flood-prone areas using remote
sensing techniques. Water Resour Manag 19(4):333–347
Jain SK, Saraf AK, Goswami A, Ahmad T (2006) Flood inundation mapping using NOAA AVHRR
data. Water Resour Manag 20(6):949–959
Jensen JR (2005) An introductory digital image processing: a remote sensing perspective. Prentice
Hall, New Jersey, p 526
Kundzewicz ZW, Hirabayashi Y, Kanae S (2009) River floods in the changing climate—observations
and projections. Water Resour Manag 24(11):2633–2646
Landis JR, Koch GG (1977) The measurement of observer agreement for categorical data. Biomet-
rics 33(1):159–174
Leica Geosystems (2005) ERDAS field guide. Leica Geosystems Geospatial Imaging, Norcross
Lillesand TM, Kiefer RW, Chipman JW (2008) Remote sensing and image interpretation. Wiley,
New York
Liu YB, Smedt FD (2005) Flood modeling for complex terrain using GIS and remote sensed infor-
mation. Water Resour Manag 19(5):605–624
Machado SM, Ahmad S (2007) Flood hazard assessment of Atrato River in Colombia. Water Resour
Manag 21(3):591–609
Macleod R, Congalton R (1998) A quantitative comparison of change-detection algorithms for mon-
itoring eelgrass from remotely sensed data. Photogramm Eng Remote Sensing 64(3):207–216
Mas J (1999) Monitoring land-cover changes: a comparison of change detection techniques. Int
J Remote Sens 20(1):139–152
Mohapatra PK, Singh RD (2003) Flood management in India. Nat Hazards 28:131–143
3246 C.S. Sharma et al.

Morrison RB, White PG (1976) Monitoring flood inundation. ERTS-1 a new window on our planet.
Prof Pap–Geol Surv (US) 929:196–208
Nagler PL, Glenn EP, Hinojosa-Huerta O (2009) Synthesis of ground and remote sensing methods
for monitoring ecosystem functions in the Colorado River delta, Mexico. Remote Sens Environ
113(7):1473–1485
Owojori A, Xie H (2005) Landsat image-based land-use and land-cover changes of San Antonio,
Texas using advanced atmospheric correction and object-oriented image analysis approaches.
In: 5th International symposium on remote sensing of urban areas, Tempe, AZ
Phillips TL (1973) LARSYS version 3 users manual. Laboratory for Applications of Remote Sensing,
Purdue University, West Lafayette
Reger B, Otte A, Waldhardt R (2007) Identifying of land cover change and their physical attributes
in a marginal European landscape. Landsc Urban Plan 81:104–113
Serra P, Pons X, Sauri D (2008) Land-cover and land-use change in a Mediterranean landscape:
a spatial analysis of driving forces integrating biophysical and human factors. Appl Geogr
28(3):189–209
Shalaby A, Tateishi R (2007) Remote sensing and GIS for mapping and monitoring land cover and
land use changes in the Northwestern coastal zone of Egypt. Appl Geogr 27(1):28–41
Singh A (1989) Digital change detection techniques using remotely sensed data. Int J Remote Sens
10 (6):989–1003
Tucker CJ (1979) Red and photographic infrared linear combinations for monitoring vegetation.
Remote Sens Environ 8(2):127–150
United Nation Development Programme-UNDP (2004) Reducing disaster risk: a challenge for
development. United Nations Development Programme, Bureau for Crisis Prevention and Re-
covery, New York, p 146
Wang F (1990) Fuzzy supervised classification of remote sensing images. IEEE Trans Geosci Remote
Sens 28(2):194–201
Yuan F, Sawaya KE, Loeffelholz B, Bauer ME (2005) Land cover classification and change analysis
of the twin cities (Minnesota) metropolitan areas by multitemporal Landsat remote sensing.
Remote Sens Environ 98(2):317–328
Yu Q, Gong P, Clinton N, Biging G, Kelly M, Schirokauer D (2006) Object-based detailed vegetation
classification with airborne high spatial resolution remote sensing imagery. Photogramm Eng
Remote Sensing 72(7):799–811
Zadeh LA (1965) Fuzzy sets. Inform Control 8(3):338–353

You might also like