Classification of LISS-III images using LVQ

Classification of LISS-III images using LVQ

Mr. Anand Upadhyay1,Dr.(Mr.) Santosh Kumar Singp,Mr. Rajesh Gupta3, Mr. Dayashankar Gupta4 Assistant Professor1,Head of IT Department2,Post Graduate Students3,4

Thakur college of Science and Commerce Thakur village,Kandivali(E),Mumbai-400101

Abstract The Classification of land cover and land use satellite image is important application of remote sensing and satellite image classification. The data acquired over various parts of Mumbai region is LISS-III. In, this paper the proposed classifier is LVQ based artificial neural network which is developed using MATLAB. It is supervised classification technique applied on the LISS-III satellite images. The classification accuracy is calculated after applying the LVQ artificial neural network using the confusion matrix and Kappa co-efficient. The classification accuracy of the proposed classifier is 94.79% and the Kappa coefficient is 0.9255 respectively.

Keywords LVQ(Learning vector quantization), Remopte Sensing

, Artificial neural network,Supervised Classification

1. INTRODUCTION

   The remote sensing is widely used technology. In remote sensing there is the satellite which is far away from the earth surface and has high power camera with high resolution. The satellite camera is very powerful; it captures the images of the various parts of the earth surface. So, these particular images are used for different purposes by different organization, agencies and other research organization for their study and planning of different resources. The one of the very important application of remote sensing is land use and land cover classification in which with help of satellite images and classification method the use of land and land covers are analyzed. In, this analysis the researcher are try to find out the change in land and land cover with different types of objects

   i.e. water, human settlements, agriculture, forest and wetland etc. [13] The Classification is one of the very important applications of satellite image processing but the very difficult part of the classification is to decide the appropriate characteristics based on that the classifier is going to classify the different classes from each other. Classification is nothing more that the partition of observation space into the disjoint region. There are different classifications methods are available for the classification of land use and land cover classification but basically all the different types of classifiers are broadly divided into two categories. The first category is known as the supervised classification technique in which the classifiers uses the past data or the training set and based on that the classifier is capable to classify the unknown data into the different classes. In the second category there is no involvement of the past data or the training set here the classifier classified the data based on the similarities or based on their nearest relation within the same class. In, this classification technique the LVQ (learning vector quantization) which is one of the types of the

   artificial neural network is used to classify the LISS-III satellite images into different classes which are water, forest, mangroves and settlements. The learning vector quantization uses the supervised classification techniques to classify the data into different classes.

2. SURVEY AND FIELD STUDY

   The supervised classification techniques required the training sets of the data of the different classes so the training set data should be the actual characteristics which properly describe the specific class so for that reason we required the survey and the field work to collect the training sample. The survey and field study plays very important role in case of satellite image classification. In the survey and field study the area or region which we want to classify that survived and the analyst tries to locate these areas on the LISS-III data. These areas are known as training sites [2]. These data are used to guide or train the classifier and then classifier is able to classify the unknown data based on the train data sets. Here, the data is collected by maps, and personal experience.

3. STUDY AREA AND CHARACTERSTICS

   The Resourcesat – 1 is designed to provide multispectral, monoscopic and stereoscopic imageries of the earths surface with its advanced on-board sensors. Indian Remote Sensing (IRS-P6) satellite gives LISS-III (Linear Imaging and Self Scanning Sensor) data. Classified land use and land cover map generated from LISS-III data is used to assess the land cover by human settlement, water, forest and mangroves. The data set of Mumbai region acquired for this research was collected via IRS-P6 resourcesat-1 satellites using LISS-III sensors in the multispectral mode by NRSA, Hyderabad, Andhra Pradesh (A.P), India. The LISS-III is a multi-spectral camera operating in four spectral bands. The LISS-III data sets consist of four different bands which are 0.52-0.59 microns (B2), 0.62-0.68 microns (B3), 0.77-0.86 microns (B4) and 1.55-1.70 microns (B5) of images. There are three bands in the visible and one near infrared. LISS-III provides data with a spatial resolution of 23.5 m unlike in IRS-P6 (where the spatial resolution is 70.5 m).The LISS-III data sets are widely used for Land use/ Land cover, Urban planning, biodiversity characterization, Forest survey, wet land mapping, environmental impact, crop acreage and production estimation of major crops, drought monitoring and assessment based on vegetation condition, snowmelt run off estimation and so on [3].

4. ARTIFICIAL NEURAL NETWORK

   The human brain consists of millions of biological neurons

   IF (The input x and associated weight wI(x) have the different class labels) THEN

   {

   which helps human to understand the pattern and based on the pattern the human can remember the different things. So, an artificial neural network (ANN) is biologically inspired from the human biological neurons. The artificial neural network often called as neural network. The artificial neural network

   Move them apart by wI (x)

   } ELSE

   {

   (t) = (t) (x wI (x) (t)).

   is a mathematical model or computational model based on biological neural networks. The artificial neural network is capable to learn or understand the different types of pattern where there is no appropriate mathematical relationships are not available for differentiation purpose of different classes. The artificial neural network consists of an interconnected group of artificial neurons and processes information using a connectionist approach to computation. The artificial neural network is dynamic in the nature because in most cases an ANN is an adaptive system that changes its structure based on external or internal information that flows through the network during the learning phase [2][14].

   LVQ(Learning vector quantization)

   The LVQ stands for Learning Vector Quantization Networks which is one of the types of the artificial neural network. The LVQ neural network was created by Teuvo Kohonen. The Learning vector quantization is multinomial classifiers of a probabilistic type. The LVQ uses the supervised and competitive learning algorithm. The LVQ neural network same as like other neural network consist of the input layer, hidden layer and output layer simultaneously. In the LVQ neural network takes the input data from the input neurons and in the neurons of hidden layers it performs the clustering which is the type of the unsupervised classification but at the output layer it matches each input withthe decided class labels which is called as the codebook. So, internally the LVQ uses the clustering which will helps the LVQ to broadly and properly classify the input data to specific class. It also uses the concept of probabilistic learning with the principle of Winner Takes All which means in the hidden layer the neurons compete with each other and the neuron which has the high activation they all will survive and other will shut down.

   The basic LVQ algorithm is a straightforward method for shifting the Voronoi cell boundaries to result in better classification. It starts from the trained SOM with input vectors x={x1, x2, x3.xn} and weights/Voronoi vectors wj= {w1, w2, w3.. wn}, and uses the classification labels of the inputs to find the best classification label for each wj.

   The LVQ algorithm then checks the input classes against the Voronoi cell classes and moves the wj appropriately:

   Algorithm:

   IF (The input x and the associated weight w(x) (i.e. the weight of the winning output node I(x) have the same class label) THEN

   {

   Move them closer together by wI (x) (t) = (t) (x wI (x) (t)).

   } ELSE

   Weights wj corresponding to other input regions are left unchanged with wj (t) = 0.

   }

   Where (t) is a learning rate and which decreases with the number of iterations/epochs of training [13].

   Input layer Hidden layer Output Layer

   Fig. 1. Architecture of Learning Vector Quantization

5. METHODOLOGY

   The algorithm is developed in MATLAB R2010. It uses a Learning vector quantization artificial neural network with LVQ1 algorithm. LISS-III satellite image consist of four different bands of images. Initially, before going to extract the feature of satellite image, all four different bands are stacked together to get the one RGB image. The following diagram represents the steps involved in satellite image classification.

   Feature Extraction of pre- processed and Enhanced LISS-III

   satellite image

   LVQ

   artificial neural network Design and Training with the sample training data

   Testing the LVQ ANN

   with the testing image and perform Classific ation

   Classif ied image

   1. Classification

      The LVQ neural networks which are trained with training data sets are used to classify the desired image. The trained LVQ neural network is ready to classify the desired image. The testing can be done with a separate testing set which is created while creating training set and based on that the classification accuracy calculated using confusion matrix and Kappa coefficients.

6. DATA AQUIRED

   LISS-III data set of Indian region is freely available online. LISS-III data sets are available in four different bands. The data is acquired in April 2013.The field work is carried out in the month of July and september2013.

7. RESULT

   The Neural Network trained with sample data. After training, the testing is done using testing dataset and the accuracy

   Fig. 2. LVQ Classification Methodology

   1. Feature Extraction

      In order to classify the satellite images, the first step is feature extraction. Here the pixel based classification methodology is adopted so from each class area region the training samples are collected in the characteristics of the R (Red), G (Green) and B (Blue) pixel value of each pixel. The training pixels are collected from the different regions of the image for the specific class and that class is assigned by the unique value. Based on the features of known classes the LVQ artificial neural network is trained. During testing when the features are provided for pixels which are not having the classes then it will give the class based on their learning.

   2. LVQ Artificial Neural Network Design

   The Matlab is very powerful mathematical computational model which provides the tool for LVQ artificial neural network based classification. The LVQ artificial neural network is designed by using the LVQ neural network toolbox. The data sets are preprocessed as the Matlab toolbox required. The LVQ artificial neural network object in the Matlab toolbox expects the samples along columns and its features along rows. Our dataset has its samples along rows and its features along columns. After preprocessing of sample data the next step is to create a LVQ artificial neural network (LVQ2) that will learn to identify the classes. Since the LVQ neural network starts with random initial weights, the results will differ. The LVQ neural network is trained to avoid this randomness. A 1-hidden layer LVQ neural network is created with 4 neurons in the hidden layer. After all these setting the network is ready to be trained. The samples data set is automatically divided into training, validation and test sets. The training set is used to teach the network. Training continues as long as the network continues improving on the validation set. The test set provides a completely independent measure of network accuracy.

   assessment is done through the confusion matrix and Kappa co-efficient. The neural network is used to classify the desired image. The confusion matrix and kappa coefficient are for the data set which is tested by LVQ neural network which is trained with the training data set.

   TABLE I. CONFUSION MATRIX

   Accuracy= (546/576)*100;

   =94.79

   Classes	Water	Forest	Mang roves	settle ment	Total	Users Acc (%)
   Water	219	0	3	0	222	98.65
   Forest	1	183	13	1	198	92.42
   Mangroves	0	10	71	2	83	85.54
   settlement	0	0	0	73	73	100
   Total	220	193	87	76	546
   Producers Acc (%)	99.55	94.82	81.61	96.05		94.79%

   =:95%

   K=0.9255(Very Good)

   Fig. 3. False colour image before classification

   Fig. 4. False colour image before classification

   Fig. 5. Colour Image After Classification

   Fig. 6. Colour Image After Classification

   Fig. 7. Colour Image After Classification

8. CONCLUSION

   In this paper learning vector quantization artificial neural network has been used for classifying LISS-III satellite image. Here, from the above confusion matrix and kappa value calculated it has found that the accuracy of classifier of the image is 94.79% and Kappa coefficient 0.9255 respectively. So, it is concluded that LVQ is also one of the good classifier to classify the satellite image for the classification purpose.

9. FUTURE ENHANCEMENT

   The LVQ is one of the good classification techniques which we can use for the other classification problem. It can also success fully apply on other types of the satellite images for the classification purposes. The classification accuracy of LVQ is also compared with the other classifiers in the future.

10. ACKNOWLEGEMENT

Thanks to NRSA, Hyderabad, Andhra Pradesh (A.P), India for providing the LISS-III datasets free online.

REFERENCES

1. K Perumal and Bhaskaran.Supervised Classification Performance of Multispectral images, Journal of Computing, Volume 2, February 2010, ISSN 2151-9617.

2. V.Turkar, 2010. Polarimetric SAR image classification by using artificial neural network, ICWET, Organized by TCET, Kandivali (E), Mumbai, India, PP.48-52, ISBN: 978-1-6558-812-4.

3. Madhubala.M, S.K.Mohan Rao and Ravindra G Babu, 2010. Classification of IRS LISS- III Images by using Artificial Neural Networks, IJCA, Special Issue on RTIPPR (3):152157.

4. Aria, E.Hosseini, J.Amini and M.R.Saradjian. Back Propagation Neural Network for Classification of IRS-1D Satellite Images, Tehran University, Iran. Print.

5. Hagan, M. T., and M. Menhaj, (1994) "Training feed forward networks with the Marquardt algorithm," IEEE Transactions on Neural Networks, vol. 5, no. 6, pp. 989-993.

6. Dr. M.Seetha, G. Malini Devi and Dr. K.V.N.Sunitha, Performance Assessment of Feature Based Image Retrieval using Neural Networks ,Volume 2, Issue ICTM 2011, February 2012, ISSN Online: 2277-2677.

7. A. K. Saha, M. K. Arora, E. Csaplovics and R. P. Gupta. Land Cover Classification Using IRS LISS III Image and DEM in a Rugged Terrain: A Case Study in Himalayas Geocarto International, Vol. 20, No. 2, June 2005.

8. Pakorn Watanachaturaporn, Manoj K. Arora, and Pramod K.Varshney. Multi source Classification Using Support Vector Machines: An Empirical Comparison with Decision Tree and Neural Network Classifiers, American Society for Photogrammetry and Remote Sensing, Vol. 74, No. 2, February 2008, pp. 239246.

9. M. Seetha, K. V. N. Sunitha, and G. Malini Devi, Performance Assessment of Neural Network and K-Nearest Neighbour Classification with Random Sub windows International Journal of Machine Learning and Computing, Vol. 2, No. 6, December 2012.

10. H.R. Keshtkara, H. Azarnivand, H. Arzani, S.K. Alavipanah, F. Mellati. Land Cover Classification Using IRS-1D Data and a Decision Tree Classifier, Keshtkar et al. / DESERT 17 (2013) 137- 146.

11. Giorgos Mountrakis, Jungho Im, Caesar Ogole, 2010. Support vector machines in remote sensing: A review, ISPRS Journal of Photogrammetry and Remote Sensing.

12. D. P. Kanungo and S. Sarkar. Use of Multi-Source Data Sets for Land Use/Land Cover Classification in a Hilly Terrain for Landslide Study, Disaster & Development, Vol. 5 No. 1 April 2011:35-51.

13. John A. Bullinaria, 2007. Introduction to Neural Computation: Guest Lecture 2

14. Mr.Anand Upadhyay, Dr (Mr.).Santosh Singh and Dr. (Mrs.) Varsha Turkar, 2014.Classification of IRS LISS-III image using Artificial Neural Network, IJPRET, ICETCSE, Organized by BNCE, Pusad (M.S), India, Volume 2(8):100-108, ISSN: 2319-507X.