Scholarly article on topic 'Identification of land cover changes in the coastal area of Dakshina Kannada district, South India during the year 2004–2008'

Identification of land cover changes in the coastal area of Dakshina Kannada district, South India during the year 2004–2008 Academic research paper on "Earth and related environmental sciences"

CC BY-NC-ND
0
0
Share paper
Keywords
{"Artificial bee colony" / "Remote sensing" / Classification / Coastal / "Land cover changes"}

Abstract of research paper on Earth and related environmental sciences, author of scientific article — J. Jayanth, T. Ashok Kumar, Shivaprakash Koliwad, Sri Krishnashastry

Abstract This study investigates land cover (LC) changes in the coastal area of Dakshina Kannada district in the state of Karnataka, South India, during the years 2004–2008 as a case study. IRS P-6, Linear Imaging Self Scanning sensor (LISS-IV) satellite images were used in the present work. Classification was carried out using artificial bee colony algorithm and support vector machine (SVM) which gave a better result compared to other traditional classification techniques. The best overall classification accuracy for the study area was achieved with an ABC classifier with an OCA of 80.35% for 2004year data and OCA of 80.40% for 2008year data, whereas the OCA in SVM, for the same training set is 71.42% for 2004 data and 71.38% for 2008 data on study area 1 and the results were optimised with respect to multispectral data. In study area 2, ABC algorithm achieved an OCA of 78.17% and MLC of 62.63% which was used to check the universality of the classifier. The classification results with post-classification technique for study area 1 indicate that urbanisation in the study area has almost increased twice. During the same time there is an increase in the forest plantation, agricultural plantation and a decrease in crop land and land without scrubs, indicates rapid changes in the coastal environment.

Academic research paper on topic "Identification of land cover changes in the coastal area of Dakshina Kannada district, South India during the year 2004–2008"

The Egyptian Journal of Remote Sensing and Space Sciences (2015) xxx, xxx-xxx

HOSTED BY

.«ïtmjj.-ggs&MÊm*

ELSEVIER

National Authority for Remote Sensing and Space Sciences

The Egyptian Journal of Remote Sensing and Space

Sciences

www.elsevier.com/locate/ejrs www.sciencedirect.com

RESEARCH PAPER

Identification of land cover changes in the coastal area of Dakshina Kannada district, South India during the year 2004-2008

J. Jayantha'*, T. Ashok Kumarb, Shivaprakash Koliwadc, Sri Krishnashastry d

a Dept. of Electronics and Communication Engineering, GSSS Institute of Engineering Technology, Mysore, India

b Dept. of Electronics and Communication Engineering, PES Institute of Engineering Technology and Management, Shivamogga,

c Dept. of Electronics and Communication Engineering, Vivekananda College of Engineering Technology, Puttur (DK), India d Dept. of Electronics and Communication Engineering, Mangalore Institute of Technology, Moodabidre, Mangalore, India

Received 25 March 2015; revised 21 July 2015; accepted 2 September 2015

KEYWORDS

Artificial bee colony; Remote sensing; Classification; Coastal;

Land cover changes

Abstract This study investigates land cover (LC) changes in the coastal area of Dakshina Kannada district in the state of Karnataka, South India, during the years 2004-2008 as a case study. IRS P-6, Linear Imaging Self Scanning sensor (LISS-IV) satellite images were used in the present work. Classification was carried out using artificial bee colony algorithm and support vector machine (SVM) which gave a better result compared to other traditional classification techniques. The best overall classification accuracy for the study area was achieved with an ABC classifier with an OCA of 80.35% for 2004 year data and OCA of 80.40% for 2008 year data, whereas the OCA in SVM, for the same training set is 71.42% for 2004 data and 71.38% for 2008 data on study area 1 and the results were optimised with respect to multispectral data. In study area 2, ABC algorithm achieved an OCA of 78.17% and MLC of 62.63% which was used to check the universality of the classifier. The classification results with post-classification technique for study area 1 indicate that urbanisation in the study area has almost increased twice. During the same time there is an increase in the forest plantation, agricultural plantation and a decrease in crop land and land without scrubs, indicates rapid changes in the coastal environment.

© 2015 Authority for Remote Sensing and Space Sciences. Production and hosting by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/

4.0/).

1. Introduction

* Corresponding author.

E-mail addresses: Jayanth.j@gsss.edu.in (J. Jayanth), Ashokkumar1968@gmail.com (T. Ashok Kumar), Spksagar2006@ gmail.com (S. Koliwad).

Peer review under responsibility of National Authority for Remote Sensing and Space Sciences.

Remote sensing (RS) data, with its ability for a synoptic view, repetitive coverage with calibrated sensors to detect changes, observations at different resolutions, provides a better alternative for the monitoring, modelling and management of natural resources and cultural processes as compared to the traditional methods (Bedawi and Kamel, 2010). Hence, in the above

http://dx.doi.org/10.1016/j.ejrs.2015.09.001

1110-9823 © 2015 Authority for Remote Sensing and Space Sciences. Production and hosting by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

context, accurate image classification results are a prerequisite. The evolution of high end computing systems and the availability of data of higher resolution (spatial, spectral, radiomet-ric and temporal) have made analysts to constantly explore the image processing and data mining techniques to exploit their potential in extracting the desired information efficiently from the RS data to improve classification accuracy. Moreover, obtaining satisfactory classification accuracy over urban/ semi urban land use/land cover (LU/LC) classes, particularly in high spatial resolution images, is a present day challenge.

Land cover change has become a central and important component in current strategies for managing natural resources and monitoring environmental changes. Land cover is defined by the attributes of the earth's land surface captured in the distribution of vegetation, water, desert and ice and the immediate subsurface, including biota, soil, topography, surface and groundwater and it also includes those structures created solely by human activities such as mine exposures and settlement (Lambin et al., 2003).

Land cover change is a dynamic process taking place on the bio-physical surfaces that have taken place over a period of time and space is of enormous importance in natural resource studies. Land cover change dynamics are important elements for monitoring, evaluating, protecting and planning of earth resources. Land cover changes are the major issues and challenges for the eco-friendly and sustainable development for the economic growth of any area. With the population explosion, human activities such as deforestation, soil erosion, global warming and pollution are very harmful to the environment. This causes land use/cover changes with the demand and supply of land in different activities. Change detection in land use and land cover can be performed on a temporal scale such as a decade to assess landscape change caused due to anthropogenic activities on the land. Land use/cover change is influenced by various natural and human activity processes. In order to improve the economic condition of the area without further deteriorating the bio-environment, every bit of the available land has to be used in the most rational way. This

Figure 1 Study area: Mangalore coastal area, Karnataka, India.

requires the present and the past land use/cover data of the area.

With the invention of remote sensing for land cover mapping is a useful and detailed way to improve the selection of areas designed for agricultural, urban and/or industrial areas of a region (Selcuk et al., 2003). Application of remotely sensed data made it possible to study changes in land cover in less time, at low cost and with better accuracy (Kachhwala, 1985) in association with GIS that provides a suitable platform for data analysis, update and retrieval (Star et al., 1997; Chilar, 2000).

There are a large number of methods to examine the changes between two fixed dates, (Lu et al., 2004, 2012). Souza (2013) adopted the distinction proposed by Coppin et al. (2004) but builds on this by dividing the "multi-date transformation" category into direct comparison and post-analysis comparison.

Threshold-based methods are commonly used to distinguish changed areas from non-changed areas (Singh, 1989; Mas, 1999) but Post classification comparison performs change detection by comparing the classification maps obtained by classifying independently two remote sensing images of the same area acquired at different times (Foody and Boyd, 1999; Lu et al., 2004). Before conducting the post classification analysis, choosing land cover (LC) classes is an important aspect because the filtering process will be applied to remove isolated pixels or noise from the classification output. Then the filtered classified image will be used in a final land cover map for change analysis (Stavros Kolios et al., 2013).

Numerous methods for remote-sensing classification are grouped into supervised and unsupervised classifiers based on the training process; parametric (statistical) and non-parametric (non-statistical) classifiers based on their theoretical modelling considering the type of distribution of data (Aurelie et al., 2013); soft and hard classifiers examine only the spectral variance ignoring the spatial distribution of the pixels belonging to the classes and other artificial intelligence methods still have limitations because of the complexities of remote sensing classification.

Land use/cover dynamics are widespread, accelerating and significant processes being driven by human actions but also produce changes that impact humans (Agarwal et al., 2002). Rawat et al. (2013) used Landsat Images to identify land cover changes in the study area for the year 1990-2010 in Ramnagar town which were categorised into five different classes, viz. built-up area, vegetation, agricultural land, water bodies and sand bar using the Maximum Likelihood Technique which showed an increase of about 8.88% in buildup area and land cover categories such as vegetation and agricultural land have decreased about 9.41% and 0.69% respectively. Dwarkish et al. identified LC change detection, in the coastal zone of Dakshina Kannada (D.K.) District of Karnataka state, using multitemporal images of LISS-III data of year 1989 and 2008 were classified using maximum likelihood algorithm and encounter changes within a period and indicate that the build-up area increased by 215%.

Figure 2 Study area 2: Mangalore City of coastal belt, Karnataka, India.

Figure 3 Technical Flow chart for the analysis of the work.

Artificial bee colony (ABC) is relatively a new member which exhibits many features including, waggle dance, bee foraging, task selection, navigation system, selection of food source, collective decision making systems, which can be used to solve complex classification problems. Satisfactory results have been obtained in multi-objective environmental/economic dispatch, data clustering and medical image classification by using ABC (Pan et al., 2010). However there are very limited studies on classification using ABC, ABC is very successful in global search and can reach/opt better scope with attribute correlation when compared to traditional statistical classifiers and requires minimum understanding of the problem domain, it also does not require complex training data to follow a normal distribution. In ABC, recruit bee update itself to cope better with attribute correlation and update is directly based on performance of classification class from the knowledge of waggle dance (Xu et al., 2010; Dorigo and Stutzle, 2005). Thus ABC should have great potential in improving remote sensing classification because of the above said advantages. In this paper, an ABC algorithm is developed to classify the remote sensing images, which also optimise the classification problem by determining the position of the food source (classes) and nectar amount of food source which represents the quality of associated class.

Genetic algorithm, gives better results for classification of medium resolution images, but they are prone to over fitting the training set and derived rule set due to mutation and crossover are difficult to interpret the classes which are spatially homogeneous, i.e., barren land, degraded land, etc. (Bandyopadhyay, 2010). Even though PSO gives better results in producing higher classification accuracy but fails to update the velocity of each particle, which is complex in implementation due to the requirement of various parameter settings for their optimal performance which may mis-classify the spectrally similar classes in LISS IV data i.e., build up land and sandy area (Bedawi and Kamel, 2010). Cuckoo search searches each proportion of each individual class within a single pixel by un-mixing all available land class information in a pixel and assigning the pixel to multiple classes. But the major drawback of the cuckoo search is that it is very unstable when feature space and training areas are changed (Yang et al., 2010). Ant colony uses a sequential covering algorithm so the rules are ordered, which makes it difficult to interpret the class at the end of the list, since the list is dependent on all the previous values of pher-omone amount. ACO takes a much longer time to classify each class then non-parametric methods (Xiaoping Liu, 2010).

In this paper, an ABC algorithm is developed for the classification of remote sensing images, which optimise classification compared to other swarm intelligence techniques. Hence in the Remote sensed data our searching element is not known initially. Just like a random walker (ant, PSO, cuckoo search, etc.) we start, but at each iteration the new values (derived) help in reaching towards the final.

• Bees are very optimal well defined workers.

• They distribute the work load among themselves which

doesn't misclassify the data.

• The dancing behaviour helps in optimal design for the selection of the same class in the data.

All the above points are taken care of in the ABC algorithm. Hence the ABC is one of the promising techniques compared with other proven techniques.

2. Data and methodology

2.1. Data used and data preparation

The image data products being used in this study are of LISS-IV (Linear Imaging and Self Scanning) sensor of IRS P-6 (Indian Remote Sensing) satellite data captured on 26th December 2007 (path: 104, row: 028; 5.8 m spatial resolution) consisting of three multispectral (MS) bands recorded at green (0.52-0.59 im), red (0.62-0.68 im) and infrared (0.770.86 im) wavelengths and panchromatic images of CARTOSAT-1 captured on 7th January 2008 (path: 0531, row: 335; 2.5 m spatial resolution) were used in this study. The satellite data were procured from the National Remote

Figure 4 Flow chart for classification using ABC algorithm.

Table 1 LU/LC classification hierarchy levels and details of the training and validation sites.

Level Training sites SVM and ABC Validation sites

1. Built up land 294 144

2. crop land 213 68

3. Plantation 220 85

4. Degraded scrub land 150 34

5. Forest plantation 175 86

6. Mangroves 58 39

7. Barren land 200 86

8. Water body 99 28

9. Land with or without scrub 139 94

11. Marshy/swampy land 56 79

1. Build up land 50 16

Total 1654 759

Sensing Centre (NRSC), Hyderabad, and Karnataka State Remote Sensing Agency (KSRAC), Bangalore, India.

2.2. Study area

The study area 1 chosen for our work is the Mangalore coastal region; its geographical location co-ordinates are between 1 2°51'32"-12°57'44"N latitude and 74051'30"-74°48'01"E longitude with an elevation of approximately 0.0 m above mean sea level (AMSL). The image dimension of the study area is 1664 x 2065 pixels in MS data and 2593 x 4616 pixels in pan-sharpened data which comprise of forest plantation, crop plantation, urban area, wetlands and water (Fig. 1). The climate in this study area is relatively mild and humid in the winter month and dry and hot in the summer. Hence this area has been considered as an ideal test-bed site for the classification.

The study area 2 is considered for universality of the ABC algorithm as a classifier and the area under investigation was the campus of the National Institute of Technology Karnataka (NITK), Surathkal, and Srinivasnagar, located 22 km North of Mangalore City on the coastal belt of Karnataka State, India (Fig. 2). The campus is spread over an area of

approximately 300 acres of land between 13°00'15"-13°01'05" N latitude and 74°47'15''-74°48'02''E longitude with an elevation of approximately 0.0 m above mean sea level (AMSL). The image dimension of the study area is 1000 x 985 pixels in pan-sharpened data. It has a good mixture of spectrally overlapping classes comprising of man-made structures and natural land cover features (Fig. 2).

2.3. Data fusion

Fine spatial resolution is necessary for an accurate description of shapes, features and structures, whereas fine spectral resolution allows better discrimination between various classes in spectral space. Hence, merging of these two types of data which are complementary in nature is beneficial for deriving higher information from the image in respect of their spectral and spatial resolutions, and this has been referred to as panchromatic sharpening. In this study, we accomplished data fusion by employing commonly used pixel level RS data merging technique called Wavelet + IHS (Hong, 2008). A cubic convolution algorithm was employed during re-sampling.

2.4. Land use/cover detection and analysis

To work out the land cover classification on panchromatic fused LISS-IV data of 2.5 m spatial resolution on 2004 and 2008 year data. The ABC classification was accomplished using MATLAB software, the supervised classification based on SVM algorithm, and validation were accomplished in ENVI image processing software as shown in Fig 3. With the help of GPS, ground verification was done for doubtful areas. Based on the ground truthing, the misclassified areas were corrected using recode option in ENVI. Eleven land cover types are identified and used in this study.

2.5. Land cover change detection and analysis

For performing land cover change detection, a postclassification detection method was employed for ABC classified data for 2004 and 2008 year data. A change matrix

Table 2 Transform divergence for the year 2004 of study area 1.

Classes Water Barren Mangroves Mashy Forest Crop Plantat Degraded Land with Build up Sandy

land swanp plant land land without scrub land area

Water - 1980 1990 1990 1990 2000 1970 1990 1970 2000 1944

Barren land 1982 - 1230 1740 1964 1984 2000 1930 1450 1980 2000

Mangroves 2000 1230 - 1450 1936 1972 1820 1640 1930 1970 1999

Mashy 1990 1723 1450 - 1878 1420 1910 1723 1790 1790 1820

swanp land

Forest plant 1990 1948 1930 1870 - 1747 1940 1728 1796 1842 1868

Crop land 2000 1980 1960 1420 1423 - 1681 1123 1483 1971 1962

Plantat 1970 2000 1820 1910 1940 1681 - 1823 1946 1961 1984

Degraded 1990 1930 1640 1723 1728 1123 1823 - 1704 1914 1929

Land with 1970 1450 1930 1790 1796 1483 1946 1704 - 1978 1946

without

Build up 2000 1980 1970 1790 1842 1971 1961 1914 1978 - 1542

Sandy area 1944 2000 1999 1820 1868 1962 1984 1929 1946 1542 -

Table 3 Transform divergence of the year 2008 of study area 1.

Classes Water Barren Mangroves Mashy Forest Crop Plantat Degraded Land with Build up Sandy

land swanp land plant land land without scrub land area

Water - 2000 2000 2000 2000 2000 2000 2000 2000 1996 1900

Barren land 2000 - 1990 1974 2000 1692 1800 1273 1346 1900 2000

Mangroves 2000 1790 - 1978 1774 1800 1848 2000 2000 1946 1989

Mashy 2000 1974 1978 - 1846 1345 1946 1981 1824 1794 2000

swanp land

Forest plant 2000 2000 1784 1846 - 1648 1799 1988 1687 1900 1986

Crop land 2000 1692 1800 1345 1648 - 1940 1748 1289 1926 1999

Plantat 2000 1800 1848 1946 1799 1940 - 1848 1748 1846 2000

Degraded 2000 1273 2000 1981 1988 1748 1848 - 1834 1981 2000

Land with 2000 1346 2000 1824 1687 1289 1748 1834 - 1998 2000

without

Build up 1996 1900 1946 1794 1900 1926 1846 1981 1998 - 1749

Sandy area 1900 2000 1989 2000 1986 1999 2000 2000 2000 1749 -

Table 4 Transform divergence of 2008 year data of study area 2.

Signature name 1 2 3 4 5 6 7 8 9 10 11 12

Grass dry 1 0 1999.91 2000 1534.96 1947.31 1993.68 2000 2000 2000 2000 1970.07 1979.05

Veg_acacia 2 1999.91 0 1922.52 1999.42 1999.99 2000 2000 2000 2000 2000 1999.92 1188.12

Veg_mix 3 2000 1922,52 0 1999.99 2000 2000 2000 2000 2000 2000 2000 1936.86

rd_interior 4 1999.42 1999.99 0 1968.69 1998.53 2000 1956.49 2000 2000 1994.7 1924.3

rd_NH 5 1947.31 1999.99 2000 1968.69 0 2000 2000 2000 2000 2000 1647.27 970.38

Open_gnd 6 1993.68 2000 2000 1998.53 2000 0 2000 1999.59 1996.95 2000 2000 2000

Water_pool 7 2000 2000 2000 2000 2000 2000 0 1997.7 2000 1995.65 2000 2000

Roof_icc 8 2000 2000 2000 1956.49 2000 1999.59 1997.7 0 1926.31 2000 2000 2000

Sand 9 2000 2000 2000 2000 2000 1996.95 2000 1926.31 0 2000 2000 2000

Water_sea 10 2000 2000 2000 2000 2000 2000 1995.65 2000 2000 0 2000 2000

Roof_sheet 11 1970.07 1999.92 2000 1994.7 <¿647^22) 2000 2000 2000 2000 2000 0 1888.26

Roof_tiled 12 1979.05 (TÎ88T25 1936.86 1924.3 1970 2000 2000 2000 2000 2000 (J888l265 0

Figure 5 Influence of average number of levels on the efficiency of bees.

(Weng, 2001) was produced with the help of ENVI software. Quantitative areal data of the overall land cover changes as well as gains and losses in each category between 2004 and 2008 were then compiled.

3. Artificial bee colony

The construction of the artificial bee colony (ABC) algorithm begins with the studies of particular intelligent behaviour of

Figure 6 Influence of selection of minimum training sample per class on the performance of bees.

bees based on their specialisation, related to their morphology (Tereshko, 2000). In statistics, ABC was used to optimise the

numerical problem (Karaboga, 2005) due to their own labour division for adjustment process and gathering information. At around the same time artificial bee colony algorithm is used in image processing and data mining due to self-organisation feature approach, heuristic approach and nearest neighbour approach. ABC algorithm is being refined over many years by several authors to facilitate the feasible region of search space.

Artificial bee colony is different compared to other swarm intelligence (SI) due to the satisfaction principle which can overcome very complex behaviour as it is a simpler process. The bee system has essential components for identifying the value of the food source depending on profitability which is expressed as a fitness function through, recruitment and abandonment of the food source which are mainly dependent on

dancing, net site selection and navigation of the food source, richness and extraction of energy and communication of a forager.

• Unemployed foragers: Main task of these bees is to exploit the food source in the search field where they do not have knowledge about it. There are two types of choices for an unemployed forager:

• Scout bee: Searches for food sources spontaneously depend on the information present inside the nest by selecting 8-30% of scout bees. The mean number of scout bees averaged over condition is about 10% (Srideepa et al., 2012).

• Recruit bee: These bees gets knowledge of the food source by attending the waggle dance done by other bees.

Figure 7 Pan-sharpened multi-spectral image using SVM Classifier (2004).

• Employed foragers: Bees find and exploit the knowledge of the food source by the waggle dance and for every food source there is only one employed bee, which memorises the location of the food source, loads the nectar and returns to unload it at the food area in the hive and checks the probability of related residual amount of nectar under three conditions.

4. Proposed algorithm for classification

An artificial bee algorithm takes a different approach for land cover classification. It's behaviour breaks a complex

classification problem to a simpler decision making process depending on the greedy randomised adaptive search heuristic (GRAH). Depending on the number of features list, agents have been assigned depending on the probability of each class used in each task and evaluate their fitness function for each land cover feature. Hence this algorithm has been evaluated in this study on pan-sharpened data of 2.5 m resolution for the years 2004 and 2008.

In this study, several assumptions have been used for training and validating the instance to correctly represent the label for the object being selected for classification. Assumptions of ABC algorithm for classification are:

Figure 8 Pan-sharpened multi-spectral image using ABC Classifier (2004).

• Pixels (digital number (DN) values) in the image are represented by bees.

• Land cover features such as water body, build up land, barren, etc. are the food sources.

The proposed ABC algorithm for the classification task is expanded until every training instance is correctly classified. The four essential elements in the proposed algorithm are:

• Initialising a set of features.

• Fitness function.

• Local search strategy and prediction strategy.

• Classification strategy.

4.1. Initialising a set of features

• Let S be a training set of objects that belongs to a mixture of classes, are divided into subsets {S1, S2, S3, ... ,Sn} each belonging to one of the I classes denoted as {C1,C2, C3,.. .,Cn}, I is a set of tasks (i = 1, ..., n) for each subset.

• Each class is assigned with a set of agents J (j = 1,..., m); and checks the resource capacity of agent bj, depending on the task i assigned to the agent j ay.

• Each task i assigned to agent j depending on the value of each task Vy using a decision variable Xj. (xy = 1 if task is assigned j; 0 otherwise).

• Bee colony is constructed using the GRAPH algorithm depending on the probability function which updates for each iteration by using features of good solutions. All the agents are determined to select a new task. Repeat these steps till all the classes are tasks that are assigned to the agents.

4.2. Fitness function

Instead of measuring the nectar amount, fitness function is used for classification to examine the types of record, which contain the feature value between the lower bound and the upper bound which can be covered by the agent to avoid mis-classification. If all tasks are covered by the agent, it means the feature values have been classified in the data.

If the class of the evaluated record is equal to the predictive class then the onlooker bee has become an employed bee attending the waggle dance and assigned with the new task of the same class. Its representation is defined as below

Fitness value = TN/TN + FP * TP/TP + FN

I. True negative (TN): Classes that are not covered by feature and that don't have class in the predictive class. II. True positive (TP): Classes that are in the features and covered by predictive class.

III. False positive (FP): Number of bees (pixel) covered by class, but the class is not covered by predictive class.

IV. False negative (FN): Number of bees (pixel) not covered by the class, but the class is covered by predictive class.

Algorithm:

1. Parameter initialisation n = number of employed bees; m = number of onlooker bees; a}- = penalty parameter for initial value for agent J

2. Initialising employed bee with GRAPH algorithm employed bee in the population A. Let set of tasks assigned to agent j be

Sj = U 8j = 1, ..., m.

Assigning each task to the list of agents, Li = {1 8i. While considering the order i = 1. If all the tasks are not assigned to repeat

Selecting new agent j* from Li using the probability function that depends on the resource needed by task i and resource dependent on agent j. Choosing the agent with greater probability has a minimal value.

Pa = '

Agent j

is assigned with task i: Sj = Sy U {/}. Let

i = i +1 and if ¡es; by* > ff remove j from any list.

Repeat step C but capacity constraints should not affect the fitness function value.

D. Let r(i) = j if i 2 Sj

Algorithm:

1. Fitness function for employed bees

fit (employed bee) = ^^c^xj j=1 i=1

+ a^^ max < ^^bijxij — aj > j=1 {i=1 )

2. For each employed bee apply neighbourhood shift

if fit(Task 1) < fit(employed bee) then mployed bee = employed bee of the same task (i.e., Tp)

else fit (shift neighbourhood) < fit(employed bee) then employed bee = employed bee of the other task (i.e.,

3. Determine probabilities by using fitness function for classification

= (S 1/fitj)—1 Pi fiti

4. Fitness for each onlooker bee

• Onlooker bees are sent to the food source depending on previously determined probabilities of employed bees.

• If the fitness value of each onlooker bee is better than the fitness value of the employed bee, solution is replaced with the onlooker solution.

if (min(fit(onlooker bee) < fit(employed bee) then

employed bee = onlooker bee

5. Again repeat step 1 and step 4 until cycle = iterations.

4.3. Exchanged search strategy

When the employed bee meets the requirement and reached the maximum cycle number, it needs to move a new food source followed by local strategy. We propose a new simpler classification local strategy which is named "exchange" to replace the original local search strategy.

Algorithm:

shift (r)

1. Let S = {i|i 2 {1, ...,n}}, k =1, neighbour task = r

2. If S = U then stop; otherwise ik is eject from rk. S = S-{ik}

3. Let j* be the agent j that minimises

ckj + x max

atj\ + aikj — bj >

I \iel,r(i)=j J )

among all agents j 2 J\{r(ik)}.

4. Assign ik to j*, output r', calculate fitness (r0) above 10% in onlooker

5. If fitness (r0) = fitness (next food source) then agent j = r0

Xij = ybestj + Uij(xkj — xki)

6. k = k + 1, return to step 2

In these phases, 10% of all possible solutions, which have the lowest fitness value, are to be updated. Hence, the proposed phase only updates poor possible solutions. The poor possible solutions are mutated around the best food source. In this

phase. where Xij is the candidate solution of new food sources, ybest, j is the global best food source with jth dimension, xkj is the food sources of jth agent and ykj is the food sources of ith dimension. k is randomly chosen food sources and they are mutually exclusive. Meanwhile the parameter Uj is a control parameter that represents random numbers within [—1,1]. As poor possible solutions are mutated around the ybest possible solution, the modified poor possible solutions would be fitter. This way, the number of fit possible solutions increases with increasing generations. Now, there exists higher probability that a selected possible solution will be mutated with a fit possible solution during employed and onlooker bees phases, as fitness of every possible solution is higher in the proposed algorithm. Hence, the produced candidate solution will be fitter than the existing possible solution.

4.4. Nectar values

The aim of the classification is to set an upper bound and lower bound for a feature value which can identify the specific class from different groups which evaluates the nectar quality of

Figure 10 Pan-sharpened multi-spectral image using ABC Classifier (2008).

Table 5 Confusion matrix and conditional kappa values of the SVM classification results obtained for 2004 year for 11 classes.

Classes 1 2 3 4 5 6 7 8 9 10 11 Row total User's accuracy (%)

1 116 2 24 1 8 2 3 156 74.36

2 10 59 1 2 62 72.03

3 0 44 9 2 57 77.19

4 4 1 31 3 39 88.57

5 1 2 7 67 8 85 78.82

6 7 29 2 39 74.36

7 1 4 51 35 2 98 57.14

8 27 27 100

9 4 79 8 91 86.81

10 24 1 9 65 74 87.84

11 1 15 16 93.75

Column total 132 53 76 38 76 37 66 37 124 78 15 759

Producer's accuracy (%) 80.22 86.79 47.83 79.49 83.75 78.38 80.00 72.97 63.20 83.33 100 OCA (%) 71.83

Kappa 0.592 0.775 0.741 0.879 0.764 0.737 0.52 1.0 0.842 0.865 0.936

Class legend: (1) Built up land, (2) crop land, (3) agriculture plantation, (4) mangroves, (5) land with/without scrubs, (6) degraded land, (7) barren land, (8) sandy area, (9) marshy/swampy land, (10) forest plantation, (11) water body.

Table 6 Confusion matrix and conditional kappa values of the SVM classification results obtained for 2008 year for 11 classes.

Classes 1 2 3 4 5 6 7 8 9 10 11 Row total User's accuracy (%)

1 116 6 19 1 9 3 4 158 74.42

2 10 40 2 1 1 54 76.67

3 4 41 8 53 77.36

4 1 31 3 35 88.57

5 1 1 13 67 2 1 85 78.82

6 7 30 2 1 39 75.00

7 1 3 51 52 2 109 46.79

8 1 33 34 97.06

9 7 61 6 74 80.45

10 1 9 66 76 86.84

11 1 13 14 81.25

Column total 133 53 92 39 80 37 70 37 124 78 15 759

Producer's Accuracy 87.22 86.79 44.56 79.49 83.75 81.08 72.86 89.19 48.80 84.62 86.67 OCA 71.83

(%) (%)

Kappa 0.680 0.749 0.737 0.879 0.764 0.737 0.476 0.969 0.791 0.854 0.808

Class legend: (1) Built up land, (2) crop land, (3) agriculture plantation, (4) mangroves, (5) land with/without scrubs, (6) degraded land, (7) barren land, (8) sandy area, (9) marshy/swampy land, (10) forest plantation, (11) water body.

Table 7 Confusion matrix and conditional kappa values of the ABC classification results obtained for 2004 year for 11 classes.

Classes 1 2 3 4 5 6 7 8 9 10 11 Row total User's accuracy (%)

1 116 2 8 1 10 1 3 141 82.27

2 10 50 2 6 68 73.53

3 5 69 13 87 79.31

4 1 30 3 34 88.24

5 1 1 14 65 3 84 77.38

6 8 30 1 40 75.00

7 1 3 56 36 2 98 57.14

8 28 28 100

9 3 75 4 82 91.46

10 13 69 82 84.15

11 1 15 16 93.75

Column total 133 53 92 39 80 37 70 37 125 78 15 759

Producer's accuracy 87.22 94.34 75.00 76.92 81.25 81.08 75.71 75.68 60.00 88.46 100 OCA 80.40

(%) (%)

Kappa 0.786 0.716 0.765 0.876 0.748 0.737 0.579 0.895 0.898 0.824 0.936

Class legend: (1) Built up land, (2) crop land, (3) agriculture plantation, (4) mangroves, (5) land with/without scrubs, (6) degraded land, (7) barren land, (8) sandy area, (9) marshy/swampy land, (10) forest plantation, (11) water body.

the source here represented by Euclidean distance. After the first ejection the free task is i0 which will be treated as end of one region, to complete other tasks of the same class in a set, i0 is assigned to other agents. The next ejection move is applied to the previous agent or to the new agent. Same steps are repeated till the single class is classified and each task is assigned with a new class and the same steps are repeated again.

4.5. Performance of artificial bee colony algorithm

Once the algorithm is built from the training set, the performance is measured in error rate depending on the estimation

of each class tested. Accordingly, it is based on the class cover percentage depending on the estimation of learning accuracy. Accordingly it is based on the estimation of learning accuracy MATLAB that relies on the accuracy obtained from the training data.

Accuracy of the training set: In this method, classifier is run on the training data itself. This method is said to be optimistic depending on three conditions.

(a) Calculate the test data record which covers each class depending on the food source.

(b) Upper and lower bounds are calculated for different classes and each one is evaluated.

(c) Select class which shows stabilisation depending on the number of bees and training samples

stabilisation = (x * fitness value of class) + (b * cover percentage)

where x and b are two weighted parameters, x€[0,1] and b = (1 — x).

Cover percentage = TP /N

The following main components in this proposed algorithm are shown in Fig. 4.

5. Implementation and result

Selection of training samples is directly related to the digital number (DN) value of the class and it is the initial step for artificial bee colony (ABC) classification. A total of 2413 samples (pixels) are used to identify the land cover (LC) classes i.e., 1654 samples as the training data set (employed bee), and 759 samples for validating classes. The number of land cover and land usage classes for classification is eleven and they are listed in Table 1. For both methods the number of training sites and validation sites are the same.

For classifying the data using ABC algorithm the following parameters have been set for the implementation purpose: no

Table 8 Confusion matrix and conditional kappa values of the ABC classification results obtained for the year 2008 for 11 classes.

Classes 1 2 3 4 5 6 7 8 9 10 11 Row total User's accuracy (%)

1 116 2 8 1 10 3 4 144 80.55

2 10 50 2 6 68 73.52

3 4 68 13 85 80.00

4 1 31 2 34 91.18

5 2 1 15 65 3 86 75.58

6 7 31 1 39 77.50

7 1 3 53 29 86 61.68

8 1 31 34 97.56

9 3 81 7 88 86.17

10 12 67 79 84.81

11 1 1 15 16 93.75

Column total 133 53 92 39 80 37 70 37 125 78 15 759

Producer's accuracy (%) 87.22 94.34 73.91 79.49 81.25 83.78 75.71 75.67 64.80 85.90 100 OCA (%) 80.40

Kappa 0.766 0.716 0.773 0.907 0.728 0.764 0.579 0.792 0.835 0.831 0.936

Class legend: (1) Built up land, (2) crop land, (3) agriculture plantation, (4) mangroves, (5) land with/without scrubs, (6) degraded land, (7) barren land, (8) sandy area, (9) marshy/swampy land, (10) forest plantation, (11) water body.

Algorithm:

Fmax-maximum values of class Fmin-minimum values of the class.

Fmax-Fmin = difference between the means the range of the class.

f is the original value of the class.

1 k1 and k2 are two random values between 0 and 1. Lower bound = f-ki * (Fmax-Fmin) Upper bound = f

+ k2 * (Fmax-Fmin)

2 Let S: = *.

3 Let S = I0, stop; or randomly choose a i0 2 I\S0,

4 Let S: = S U {i0}, Job i0 is ejected from r(i0)if r' = r j* is a agent of j with

ckj + xj max

aij\ + aikj — bj >

I \i2j,r(i)=j J J

among all agents j 2 J\{r(ik)}. and l =0;

5 If B(il)\{ik\k < l} = * return to step 3 otherwise l = l +1 and proceed to step 7.

• If Il belongs to B(i1) select the following steps

• If i0 is inserted into r(i1) by r'(i0) and output r' select agent j else r'(i0):= j* new task of same class with i0 = i

6 Repeat step 4 until complete task of class is completed

7 Selection of new class repeat from step 1.

of bees = 220 (employed bees = 60 and onlookers bee = 160); minimum training samples = 20, maximum uncovered training samples = 12; and maximum iterations for onlooker bee = 220.

For the implementation of SVM Classifier, we kept gamma function value of 0.167 in the kernel function with penalty parameter value of 100 and a pyramid value of 0.

The transform divergence (TD) measure for the year 2004 data is shown in Table 2 for the year 2008 data transform divergence is shown in Table 3 for study area 1. Signature separability for each class should be in the range of 2000 maximum and 1800 minimum. Classes like built up area and sand shows serious spectral overlapping and exhibit a poor transform divergence of 1542.26 (2004 year data) and 1586.25 for the year 2008 data. Even though forest and agricultural plantation are not distinct, they still maintain an excellent separability in both the years data sets and are indicated in detail in Tables 2 and 3. Barren land and mangroves show a spectral overlapping of very low transform divergence value of 1230.96 (for 2004 year data) and showed an excellent separa-

bility in 2008 data (TD-1790.25). Tables 2 and 3 show severe spectral overlapping between the following classes like built-up versus sand, wetlands versus mangroves, degraded scrub land versus land with or without scrubs, and forest versus agricultural plantation.

For study area 2, it is found from the obtained TD (Table 4) that for all the three bands, there is a severe spectral overlapping of classes like veg_mix-veg_acacia, rd_interior-grass_dry, rd_NH-grass_dry, sand-roof_rcc, roof_sheet-rd_NH, roof_ sheet-grass_dry, roof_tiled-veg_acacia, roof_tiled-veg_mix, roof_tiled-rd_interior, roof_tiled-grass_dry and roof_tiled-roof_sheet which resulted in multimodal distribution of histograms. The worst affected classes are circled on the table, and it is expected that the classifiers will show a reduced accuracy for the spectrally overlapping classes.

In the training data set, these characteristics of ABC would make it a very viable classifier because one of the most difficult aspects of supervised classification is the selection of a sufficient number of representative training sites for each class. Much time and expense can be saved if fewer samples suffice

Table 9 Confusion matrix and conditional kappa values of the classification results obtained for SVM for 11 classes (2004).

Classes 1 2 3 4 5 6 7 8 9 10 11 Row total UA ('

1 18 18 100

2 56 3 1 3 63 88.89

3 14 1 15 93.33

4 31 78 109 28.44

5 65 9 40 114 57.02

6 12 1 75 2 90 83.33

7 7 1 116 1 36 1 162 71.60

8 4 32 2 36 88.89

9 16 16 100

10 2 3 1 21 1 3 147 178 82.58

11 1 1 21 10 41 74 55.41

Column total 18 66 21 43 68 178 147 34 19 239 42 791

PA (%) 100 84.85 66.67 72.09 95.59 42.13 78.91 94.12 84.21 61.51 97.62 OCA

Kappa 1.0 0.87 0.93 0.24 0.53 0.79 0.65 0.88 1.0 0.76 0.59

Class legend: (1) Water body, (2) barren land, (3) degraded land, (4) mangroves, (5) land with/without scrubs, (6) land with/without scrubs, (7) built-up land, (8) agricultural plantation, (9) marshy/swampy land, (10) forest plantation, (11) crop land.

Table 10 Confusion matrix and conditional kappa values of the classification results obtained for ABC for 11 classes.

Classes 1 2 3 4 5 6 7 8 9 10 11 Row total UA (

1 18 18 100

2 53 3 2 58 91.38

3 14 1 15 93.33

4 31 17 48 64.58

5 46 17 18 81 56.79

6 5 12 139 1 157 88.54

7 8 2 96 1 25 132 72.73

8 5 1 33 39 84.62

9 17 17 100

10 1 2 20 30 2 182 1 237 76.79

11 1 22 11 40 73 54.79

Column total 18 66 21 43 68 178 147 34 19 239 42 791

PA (%) 100 80.30 66.67 72.09 67.65 78.09 65.31 97.06 100 76.15 95.42 OCA

Kappa 1.0 0.90 0.93 0.62 0.53 0.85 0.67 0.83 1.0 0.68 0.52

Class legend: (1) Water body, (2) barren land, (3) degraded land, (4) mangroves, (5) land with/without scrubs, (6) land with/without scrubs, (7) built-up land, (8) agricultural plantation, (9) marshy/swampy land, (10) forest plantation, (11) crop land.

Table 11 Confusion matrix and conditional kappa values of the classification results obtained for SVM for 11 classes.

Classes 1 2 3 4 5 6 7 8 9 10 11 Row total UA ('

1 18 18 100

2 57 2 2 2 63 90.48

3 15 15 100

4 32 46 78 41.03

5 1 54 16 71 76.06

6 3 11 116 3 2 135 85.93

7 3 6 102 1 37 1 150 68.00

8 4 2 33 2 41 80.49

9 19 19 100

10 1 2 7 2 41 169 1 223 75.58

11 1 1 1 12 10 38 62 61.29

Column total 18 66 21 43 68 178 147 34 19 239 42 791

PA (%) 100 86.36 71.43 74.42 79.41 65.71 69.39 97.06 100 70.71 90.48 OCA

Kappa 1.0 0.89 0.87 0.37 0.74 0.82 0.61 0.79 1.0 0.66 0.59

Class legend: (1) Water body, (2) barren land, (3) degraded land, (4) mangroves, (5) land with/without scrubs, (6) land with/without scrubs, built-up land, (8) agricultural plantation, (9) marshy/swampy land, (10) forest plantation, (11) crop land.

to the need of the classifier to produce fairly good classification accuracy. However, in general, it is inferred that larger dataset size is favourable in improving the learning accuracy and the overall classification accuracy in the ABC classifier.

From Fig. 5, we notice that the classification accuracy is better in the interval 3.4-13.2 in this range the classification accuracy ranges from a minimum of 72-98%.

Fig 6 depicts convergence of accuracy with training and validation sites. When the number of training samples per class is minimum there is an increase in the convergence behaviour of the bee and produces a better accuracy in the defined class. Total time taken by the ABC is 5 min to complete the identification of each class.

The study was carried out to compare the performance of MLC and ABC classifier as shown in Figs. 7-10, the number of classes considered is eleven and the data of the years 2004

and 2008 are considered. The classification results with confusion matrix of SVM are tabulated in Tables 5 and 6 and for ABC in Tables 7 and 8. The highest OCA of 80.35% is obtained in ABC for 2004 year data and OCA of 80.40%, whereas the OCA in SVM, for the same training set is 71.42% for 2004 data and 71.38% for 2008 data. For 2004 year data, ABC exhibits an improvement of 7% in PA (87.22%) and 7.91% in UA (82.27%) when compared over SVM (PA: 80.22% and UA: 74.36%) for Build up Land. Further, for agricultural Plantation land, the ABC has recorded a mark improvement of 27.17% in PA (74.21%) over SVM (44.56) with no appreciable difference in UA. ABC has performed marginally better over SVM on crop land, barren land, degraded or scrub land and marshy/swampy land with 3.4% and 2.15% lesser PA and UA, respectively. Hence, in the both the years (2004 and 2008), the ABC is able to maintain approximately 8.99% higher OCA. The kappa, in turn,

Table 12 Confusion matrix and conditional kappa values of the classification results obtained for artificial bee colony algorithm for 11 classes with MS (2008).

Classes 1 2 3 4 5 6 7 8 9 10 11 Row total UA (%)

1 18 18 100

2 56 2 3 1 2 64 87.50

3 15 15 100

4 26 28 54 48.15

5 1 56 1 10 68 82.35

6 3 17 132 2 2 156 84.62

7 5 6 108 1 32 1 153 70.59

8 4 31 35 88.57

9 19 19 100

10 1 2 6 1 35 1 184 230 80.00

11 1 15 9 39 63 61.90

Column total 18 66 21 44 68 178 147 34 19 239 42 791

PA (%) 100 84.85 71.43 60.47 82.35 74.16 73.47 91.18 100 76.99 92.86 OCA 78.17

Kappa 1.0 0.864 1.0 0.454 0.808 0.806 0.646 0.881 1.0 0.724 0.599

Class legend: (1) Water body, (2) barren land, (3) degraded land, (4) mangroves, (5) land with/without scrubs, (6) land with/without scrubs, (7) built-up land, (8) agricultural plantation, (9) marshy/swampy land, (10) forest plantation, (11) crop land.

is improved from 0.026 to 0.084; approximately double. Finally, the class mangroves and sand area is benefited as it could record an increase in kappa from 0.72 to 0.864 on comparison with 11 classes. Hence, it is evident from the above that ABC has never the accuracy to go down even when the classes are spectrally overlapping and spatially non-homogeneous. For all the 11 classes, the ABC has maintained a higher kappa value over SVM.

Hence, in the both the years (2004 and 2008), the ABC is able to maintain approximately 8.99% higher OCA. The kappa, in turn, is improved from 0.026 to 0.084; approximately double. Finally, the class mangroves and sand area are benefited as it could record an increase in kappa from 0.72 to 0.864 on comparison with 11 classes. Hence, it is evident from the above that ABC has never let the accuracy go down even when the classes are spectrally overlapping and spatially non-homogeneous. For all the 11 classes, the ABC has maintained a higher kappa value over SVM.

The classwise comparison of the PA, UA and Kappa obtained in SVM and ABC for 11 LC features which is optimised using Multi spectral (MS) data is given in Tables 9 and 10 for 2004 year data and Tables 11 and 12 for 2008 year data. The classified images of SVM and ABC with the original LISS-IV MS data (5.8 m) are shown in Figs. 11-14.

The results in Tables 10 and 12 show that the highest OCA of 78.17% for 2008 year data and 76.16% for 2004 year data are obtained in the ABC algorithm with an improvement of 7% from 2004 year data and 9% from 2008 year data over SVM. Hence it is evident from the comparison that the ABC algorithm never let the accuracy go down and maintained higher Kappa values even with the multispectral data which are spectrally overlapping. But the most undesirable effect seen in the SVM classified image is that the pixels corresponding to the class. Built-up land has spread all over the image at the borders of class land with/without scrubs making it unrealistic (Figs. 11-13 and the corresponding regions in Figs. 12-14),

whereas the ABC classified images are free from this effect. In comparison with PAN sharpened image and Multispectral image ABC algorithm maintained higher OCA when compared with the SVM Classifier.

The study area 2 was carried out on a semi-urban area to compare the performance of both the classifiers at 11 classes. The classification results are tabulated in Tables 13 and 14 shows the performance of MLC and ABC classifiers as shown in Fig. 15.

The highest OCA of 78.17% is obtained in ABC for the training dataset size of 3090 pixels, whereas the OCA in SVM, for the same training set is 62.63%. ABC exhibits an improvement of 39.33% in PA (74.16%) and 0.84% in UA (84.62%) over SVM (PA: 34.83% and UA: 83.78%) for class veg_mix. Further, for class rcc_sheet, the DTC has recorded a marked improvement of 23.85% in PA over SVM with a marginal appreciable difference in the corresponding UA of 5.73%. ABC has performed marginally better over SVM on

Acacia with 27.9% lesser PA and 20.81% increase in UA, respectively. The class Int_road (Interior) has experienced a trade-off in its UA and PA between both the classifiers.

Comparison of the classification results of SVM and ABC (for study area 1) with official census data is shown in Table 15. When these results are compared, ABC algorithm matches nearer with the census data. Census data may contain errors due to tax reasons and inherent errors.

On the other hand comparison of the percentages of land is depicted in Fig 16. Degraded land and built up land have increased by 10.26% and 17.56% respectively. There is also an increase in 14.32% Forest plantation due to social forestry activities taken by the state forest department. The area under crop land was decreased by 8.89% due to conversion of crop land to commercial use along the major highway due to an increase in state highway 66. The area of mangroves and land with/without scrubs has been reduced by 8.2% respectively.

Figure 14 ABC algorithm for Multispectral image for 2008 year data.

Figure 13 SVM Classifier for Multispectral image for 2008 year data.

,0> H>

Table 13 Confusion matrix and conditional kappa values of the classification results obtained for SVM for 11 classes (study area 2).

Classes 1 2 3 4 5 6 7 8 9 10 11 Row Total UA%

1 18 18 100

2 55 12 7 74 74.32

3 15 2 17 88.24

4 38 101 139 27.34

5 58 9 53 120 48.33

6 5 4 62 1 2 74 83.78

7 4 3 84 1 32 1 125 67.20

8 4 10 33 4 51 64.71

9 19 19 100

10 2 2 7 1 32 127 171 74.21

11 1 14 13 39 67 58.21

Column 18 66 21 43 68 178 147 34 19 239 42 875

PA% 100 83.33 71.43 88.37 85.29 34.83 57.14 97.06 100 53.14 92.86 OCA

Kappa 1.0 0.72 0.87 0.23 0.43 0.79 0.60 0.63 1.0 0.64 0.599

Class legend: (1) Pool, (2) grass dry, (3) sand, (4) Acacia, (5) NH, (6) Veg_mix, (7) int_road, (8) open_gnd, (9) Sea, (10) rcc_ sheet, (11) tiled_roof.

Table 14 Confusion matrix and conditional kappa values of the classification results obtained for artificial bee colony for 11 classes (study area 2).

Classes 1 2 3 4 5 6 7 8 9 10 11 Row UA

total (%)

1 18 18 100

2 56 2 3 1 2 64 87.50

3 15 15 100

4 26 28 54 48.15

5 1 56 1 10 68 82.35

6 3 17 132 2 2 156 84.62

7 5 6 108 1 32 1 153 70.59

8 4 31 35 88.57

9 19 19 100

10 1 2 6 1 35 1 184 230 80.00

11 15 9 39 63 61.90

Column 18 66 21 43 68 178 147 34 19 239 42 875

PA (%) 100 84.85 71.43 60.47 82.35 74.16 73.47 91.18 100 76.99 92.86 OCA

Class legend: (1) Pool, (2) grass dry, (3) sand, (4) Acacia, (5) NH, (6) Veg_mix, (7) int_road, (8) open_gnd, (9) Sea, (10) rcc_sheet, (11) tiled_roof.

Figure 15 Subset of classified images.

Table 15 LC changes in Dakshina Kannada according to our results and census data.

Land cover categories Classification results (HA) Census data

MLC ABC

2004 2008 Difference 2004 2008 Difference 2004 2008 Difference

Water bodies 1510.259 1717.558 207.299 1313.23 1435.23 121.998 1417.74 1546.394 134.56

Fallow land 521.574 383.067 -138.507 490.093 328.064 -162.029 505.83 355.56 -150.27

Mangroves 348.599 395.706 47.107 417.741 235.851 -181.890 404.17 285.77 -114.40

Marshy/swampy land 344.727 303.153 -41.575 646.41 300.136 -346.274 542.26 356.25 -315.56

Crop land 939.721 566.950 372.771 740.462 612.62 -127.842 700.25 582.23 -118.02

Forest plantation 595.488 694.144 98.656 584.08 779.178 195.098 600.23 821.29 221.06

Agricultural plantation 390.936 448.613 57.677 367.387 618.256 250.869 349.85 539.56 189.70

Degraded land 576.569 666.718 90.149 406.3 552.61 146.310 481.44 609.6 128.16

Land with/without scrubs 614.443 396.988 -217.455 315.04 243.054 -71.986 411.36 326.25 -85.11

Built-up-land 1043.002 1558.703 515.701 1493.39 2129.53 636.133 1365.23 2045.32 680.23

Sandy area 140.442 152.402 11.960 146.62 126.256 -20.373 142.36 115.36 -27.04

The sandy area of 7.23% was converted to build up land due to the creation of a port near the kulur district. The agricultural plantation was increased by 15.23% respectively due to the conversion of crop land and barren land to plantation. The marshy/ swampy areas were decreased by 50% due to large-scale filling in baikampady and kottur due to expansion of build up areas.

To understand land encroachment in different land categories, a change detection matrix (Table 16) was prepared

which reveals changes that occurred during the last two decades (2004-2008) which are shown in Table 17.

5.1. Discussion

Coastal region of the DK district of Karnataka is experiencing significant and contiguous LC changes mainly due to urbanisation and intensification of agriculture and are of

x/ S ^ W J* J* <$>

Jf ^ J"

¿f J- J

Classes

Figure 16 The percentage of each class was compared with ABC classified image at the dates of 2004 and 2008.

greater importance in local ecosystems, are mainly stressed by increasing human activities and tourist activity.

As far as we know, our study comprises a first effort to study LC changes using satellite remote sensing. There are several studies in which area under conventional cultivation, such as rice paddies has decreased by 8%, but the area under agricultural plantations such as coconut plantation and areca nut plantation remains constant and the area of mangroves was decreased by 6% whereas there is a 2-fold increase in the forest plantation due to social forestry. The area covered by water remained almost constant at 20 km2.

The study area has degradation in barren land of 10% respectively during the study period and also agricultural lands were used for commercial use for expansion all along the major highways and construction of seven grade separators due to an increase in vehicular traffic, has increased the buildup area. The major driving force for fast urbanisation during the year

is the renovation of a new broad gauge railway line connecting Mangalore to Surathkal, upgrading Mangalore city corporation, expansion of Mangalore university at mangalagangotri, setting up of SEZ IT and IT enabling services and the industrial estate at baikampady has shown an increase in build up area. National highway 66 was increased to four lanes and road widening works taken up in Mangalore City have contributed to the increase in urban areas.

6. Conclusions

The study demonstrated the effect of integration of spectral and spatial features in classification of coastal areas which often display fragmented, heterogeneous land cover features on panchromatic sharpened LISS-IV data of 2.5 m spatial resolution for 11 Land cover classes which are arising from absent or ineffective local planning. The findings of this study suggest that ABC can perform better results in classification of remote sensing in the Mangalore coastal area, India. The comparison of classification results is carried out between the ABC and SVM method using OCA. OCA of 80.35% is obtained in ABC for 2004 year data and OCA of 80.40%, whereas the OCA in SVM, for the same training set is 71.42% for 2004 data and 71.38% for 2008 data. Compared to SVM, the ABC tends to be more effective for the classification of RS data.

From an environmental point of view, spatial changes in the class are mainly due to the residential development and the improvement/expansion of the road network by almost twice during the study period and the size of the barren land is increased up to 10%. During the examined period the cultivated land such as rice paddies was decreased upto 20% but the area under agricultural plantation such as coconut and arecanut plantation so therefore total area under agriculture remained same. The areas of mangroves are decreased by 6% respectively, whereas there is an increase by 2 times of forest plantation due to social forestry and forest plantation activities taken by the state forest department. The marshy/swampy areas are decreased by 23% due to filling up of marshy land for expansion of urban build up areas.

Table 16 Land cover conversion matrix (%) between the initial state (2004) and the final state (2008) using ABC algorithm.

L2008 Water Barren Mangroves Marshy Crop Forest Plantation Degraded Land with Buildup Sand

land scrubs

Water 97.26 0.000 0 0 0 0 0 0 0 0 0.021

Barren land 0 66.581 0.026 1.145 5.689 3.986 14.586 0.108 16.758 0.293 0.093

Mangroves 0.03 0.027 97.191 0.023 0.056 0.026 0.093 0.29 0.092 0.084 0.446

Marshy/swampy 1.12 0.032 0.112 65.589 0.089 0.152 1.145 0.095 1.002 1.468 0

Crop land 1.59 0.041 0.002 0.258 75.125 0.08 2.593 1.145 2.589 1.699 0

Forest plantation 0 22.268 0.732 1.235 0.986 89.258 0.135 24.258 12.258 0.742 1.415

Agricultural 0 3.087 0.043 0.589 6.568 2.568 70.568 1.316 5.365 0.142 0

plantation

Degraded scrub 0 0.247 0.024 1.02 1.589 1.235 0.29 67.958 0.258 1.348 0.172

Land with scrub 0 0.057 1.886 0.986 1.256 0.986 2.699 0.02 59.256 0.434 0.235

Build up land 0 7.642 0.008 29.13 8.658 0.835 7.235 4.694 1.589 90.222 3.558

Sandy area 0 0.021 0.018 0.025 0.036 1.235 0.201 0.049 0.358 3.568 94.42

Total 100 100 100 100 100 100 100 100 100 100 100

Table 17 Spatial Changes between the initial state (2004) and the final state (2008) using ABC algorithm.

1. Water body - all areas of water, Irrigated lands

2. Barren land - areas containing silt, clay, or with little or no "green" vegetation present which are widely spaced in the green vegetated categories and crop land is temporarily barren

3. Mangroves - areas where forest or shrub land vegetation accounts for 30-100% of the cover and are covered with water

4. Marshy/swampy land - areas are covered with grass and water

5. Crop plantation - areas used for production of crops and graminoid crops was 75-100% of the cover

6. Forest plantation - areas are dominated by tree cover (natural/semi-natural) where 75% or more and respond to seasonal changes

7. Degraded land - areas are dominated by grasses and areas of sparse vegetation cover converting from one land cover to another

8. Agricultural plantation - areas containing coconut plantation, arecanut and vineyard plantation of 50%

9. Land with/without scrubs - areas dominated by shrub which are generally greater than 25% of the cover, when the cover is with other life forms

10. Build up land - includes areas with a mixture of constructed materials that account for 30-80% and also includes commercial/ industrial/transportation

11. Sandy area

Water body almost remained constant, a small difference of 2.18% of changes was noticed. When recognising the spatial changes, there was a partial dispersion considered due to a change in the water level of the Nethravati river, and also a change in sea level due to the water flow and climatic condition. Swampy land and agricultural areas constitute 1.12% and 1.59% of spatial class changes at pixel basis Barren land in 2004 was changed to the classes of forest plantation of 22.268% and build up land 7.642% respectively. But 66.581% of barren land remains unchanged. Considering the spatial changes from one class to another class 3.087% of spatial changes is estimated due to probable irrigation in agriculture area changed the moisture level of land surface. Mangroves, degraded scrub land and land with/without scrubs was changed may be due to spectral signature and misleading the classifier

Mangroves almost remained constant, a small difference of 2% of changes was noticed. Due to spectral signature misleading the classifier with land with scrubs class

Regarding the class of Marshy/swampy land it remained constant of 65.589% respectively, but due to driving forces in industrial site in the baikampadi area and due to rapid urbanisation the spatial changes among the class showed an increase of 29.13% utilisation of marshy/ swampy land. Marshy/swampy land was converted to barren land and degraded scrub land due to cover in soil area of 1.145% Total amount of change among the classes from the 2004-2008. More specifically crop land was changed to build up land with 8.65% due to construction of road and consequently the crop land class was converted to agriculture plantation with 6.538%. Some regions are converted to barren land temporarily due to a delay in production of crops, spatial change specified with changes in 5.689% from crop land to barren land and also to land with/without scrubs The number of pixels that were classified as Forest plantation remained the same in 2008 at the percentage of 89.258%. Spatial changes with the mixed response of agricultural plantation and degraded scrub land near the shoreline caused the misclassification As a consequence there is a change of degraded land to forest plantation and conversion to a barren land with a percentage of 24.25% and 4.69%

Agriculture plantation remained the same in the final state at the percentage of 70.568%. The most important change in this class was the conversion to the build up land to 7.235% and due to degradation some of the land was converted to Barren land with an improvement of 14.586

Areas are dominated by shrub due to degradation with the land cover features they are converted to barren land with 16.578%. and areas were converted to forest plantation by the state forestry department by 12%

The buildup class remained constant at 90.22% and some of the areas with vegetation and barren land were converted to Build up. Sand area was misclassified due to low spectral distance in the classes Areas remain constant at 94.42%. Due to construction of port in the area sand area was converted to Build up

Examination of visual inspection of the classified images revealed that ABC misclassified fewer pixels as Mangroves and Barren land in the southern road confluence than either SVM. This suggests that slight improvement is needed with the Onlooker bees which creates a large number of boxes in feature space which misguides the class separation in mixed classes when the threshold is reached to a maximum limit. In future research classification rule

can be applied using XOR condition to identify the classes.

Acknowledgments

The authors would like to graciously thank Dr. V.P. Laksh-mikanth, Karnataka state remote sensing application center, Bangalore for providing the data products for the study.

Thanks also go to the Dr. Dwarkish Associate professor, NITK surathkal for providing the necessary software support for this study. The authors would finally like to thank colleagues involved at the Karnataka state remote sensing application center (KSRAC), Mysore, for their assistance in field work and other contributions to this study.

References

Aurelie, V., Vladimir, A., Moser, G., 2013. Supervised classification of multi-sensor and multi-resolution remote sensing images with a hierarchical copula-based approach. IEEE Trans. Geosci. Remote Sens. 20 (3), 1-18. Bedawi, N., Kamel, M., 2010. Segmentation of very high resolution remote sensing imagery of urban areas using particle swarm optimization algorithm. In: Image Analysis and Recognition: Lecture Notes in Computer Science. Springer-Verlag, New York.

Dorigo, M., Stutzle, T., 2005. Ant Colony Optimization. Prentice-Hall, ND.

Lambin, E.F., Geist, H.J., Lepers, E., 2003. Dynamics of land use and land cover change in tropical regions. Annu. Rev. Environ. Resour. 28, 205-241.

Pan, Q.K., Fatih, M., Suganthan, P.N., Chua, T.J., 2010. A discrete artificial bee colony algorithm for the lot-streaming flow shop scheduling problem. Inf. Sci. 26, 65-78. Srideepa, B., Bharadwaj, A., Gupta, D., Panchal, V.K., 2012. Remote sensing image classification using artificial Bee colony algorithm. Int. J. Comput. Sci. Inf. 2 (3), 67-71. Tereshko, V., 2000. Reaction diffusion model of a honeybee colony's foraging behavior. In: Parallel Problem Solving from Nature VI. Springer-Verlag, New York. Xu, C., Duan, H., Liu, F., 2010. Chaotic artificial bee colony approach to uninhabited combat air vehicle (UCAV) path planning. Aerosp. Sci. Technol. 14, 535-541.