Artificial Neural Network: A Tool for Classification of Land Use and Land Covers Using Satellite Images

Similar documents
Part 8: Neural Networks

Artificial Neural Network Approach for Land Cover Classification of Fused Hyperspectral and Lidar Data

EFFECT OF ANCILLARY DATA ON THE PERFORMANCE OF LAND COVER CLASSIFICATION USING A NEURAL NETWORK MODEL. Duong Dang KHOI.

Artificial Neural Network

A Feature Based Neural Network Model for Weather Forecasting

ARTIFICIAL NEURAL NETWORK PART I HANIEH BORHANAZAD

Deriving Uncertainty of Area Estimates from Satellite Imagery using Fuzzy Land-cover Classification

A Novel Activity Detection Method

Keywords- Source coding, Huffman encoding, Artificial neural network, Multilayer perceptron, Backpropagation algorithm

Neural Networks Introduction

Artificial Neural Network Method of Rock Mass Blastability Classification

1. Introduction. S.S. Patil 1, Sachidananda 1, U.B. Angadi 2, and D.K. Prabhuraj 3

Address for Correspondence

Neural Networks and the Back-propagation Algorithm

Introduction Biologically Motivated Crude Model Backpropagation

COMPARING PERFORMANCE OF NEURAL NETWORKS RECOGNIZING MACHINE GENERATED CHARACTERS

Artificial Neural Networks. Historical description

Classification of High Spatial Resolution Remote Sensing Images Based on Decision Fusion

EEL 851: Biometrics. An Overview of Statistical Pattern Recognition EEL 851 1

STUDY OF NORMALIZED DIFFERENCE BUILT-UP (NDBI) INDEX IN AUTOMATICALLY MAPPING URBAN AREAS FROM LANDSAT TM IMAGERY

(Feed-Forward) Neural Networks Dr. Hajira Jabeen, Prof. Jens Lehmann

Artificial Neural Network and Fuzzy Logic

URBAN LAND COVER AND LAND USE CLASSIFICATION USING HIGH SPATIAL RESOLUTION IMAGES AND SPATIAL METRICS

Artificial Neural Networks Examination, June 2005

Machine Learning. Neural Networks. (slides from Domingos, Pardo, others)

A Method to Improve the Accuracy of Remote Sensing Data Classification by Exploiting the Multi-Scale Properties in the Scene

Application of Artificial Neural Networks in Evaluation and Identification of Electrical Loss in Transformers According to the Energy Consumption

EE04 804(B) Soft Computing Ver. 1.2 Class 2. Neural Networks - I Feb 23, Sasidharan Sreedharan

Introduction To Artificial Neural Networks

Supervised Learning. George Konidaris

ECE662: Pattern Recognition and Decision Making Processes: HW TWO

Combination of M-Estimators and Neural Network Model to Analyze Inside/Outside Bark Tree Diameters

Artificial Neural Networks

Machine Learning. Neural Networks. (slides from Domingos, Pardo, others)

STA 414/2104: Lecture 8

Data Mining Part 5. Prediction

An artificial neural networks (ANNs) model is a functional abstraction of the

Neural Networks and Ensemble Methods for Classification

Artificial Neural Networks. Edward Gatt

This is trial version

Artificial Neural Networks Examination, June 2004

Unit 8: Introduction to neural networks. Perceptrons

Machine Learning. Neural Networks. (slides from Domingos, Pardo, others)

ARTIFICIAL NEURAL NETWORKS گروه مطالعاتي 17 بهار 92

Selection of Classifiers based on Multiple Classifier Behaviour

Neural Networks biological neuron artificial neuron 1

Digital Trimulus Color Image Enhancing and Quantitative Information Measuring

Short Term Load Forecasting Based Artificial Neural Network

Simple Neural Nets For Pattern Classification

KNOWLEDGE-BASED CLASSIFICATION OF LAND COVER FOR THE QUALITY ASSESSEMENT OF GIS DATABASE. Israel -

AI Programming CS F-20 Neural Networks

Lecture 4: Feed Forward Neural Networks

Study of a neural network-based system for stability augmentation of an airplane

Machine Learning to Automatically Detect Human Development from Satellite Imagery

A FUZZY NEURAL NETWORK MODEL FOR FORECASTING STOCK PRICE

IMPROVING REMOTE SENSING-DERIVED LAND USE/LAND COVER CLASSIFICATION WITH THE AID OF SPATIAL INFORMATION

Machine Learning. Neural Networks

Principals and Elements of Image Interpretation

Active Sonar Target Classification Using Classifier Ensembles

Mining Classification Knowledge

Neural Networks: Introduction

Remote Sensing and GIS Techniques for Monitoring Industrial Wastes for Baghdad City

Introduction to Natural Computation. Lecture 9. Multilayer Perceptrons and Backpropagation. Peter Lewis

Hand Written Digit Recognition using Kalman Filter

HIERARCHICAL IMAGE OBJECT-BASED STRUCTURAL ANALYSIS TOWARD URBAN LAND USE CLASSIFICATION USING HIGH-RESOLUTION IMAGERY AND AIRBORNE LIDAR DATA

M.C.PALIWAL. Department of Civil Engineering NATIONAL INSTITUTE OF TECHNICAL TEACHERS TRAINING & RESEARCH, BHOPAL (M.P.), INDIA

Neural Networks and Fuzzy Logic Rajendra Dept.of CSE ASCET

SOIL MOISTURE MODELING USING ARTIFICIAL NEURAL NETWORKS

Fundamentals of Photographic Interpretation

How to do backpropagation in a brain

Land cover/land use mapping and cha Mongolian plateau using remote sens. Title. Author(s) Bagan, Hasi; Yamagata, Yoshiki. Citation Japan.

Artificial Intelligence (AI) Common AI Methods. Training. Signals to Perceptrons. Artificial Neural Networks (ANN) Artificial Intelligence

STA 414/2104: Lecture 8

A Neuro-Fuzzy Scheme for Integrated Input Fuzzy Set Selection and Optimal Fuzzy Rule Generation for Classification

INTERNATIONAL JOURNAL OF GEOMATICS AND GEOSCIENCES Volume 2, No 3, 2012

Artificial Neural Networks (ANN)

Urban Tree Canopy Assessment Purcellville, Virginia

Abstract. TECHNOFAME- A Journal of Multidisciplinary Advance Research. Vol.2 No. 2, (2013) Received: Feb.2013; Accepted Oct.

INTRODUCTION TO NEURAL NETWORKS

Non-linear Measure Based Process Monitoring and Fault Diagnosis

Urban land cover and land use extraction from Very High Resolution remote sensing imagery

EMPIRICAL ESTIMATION OF VEGETATION PARAMETERS USING MULTISENSOR DATA FUSION

Introduction to Artificial Neural Networks

AN INTRODUCTION TO NEURAL NETWORKS. Scott Kuindersma November 12, 2009

Lecture 4: Perceptrons and Multilayer Perceptrons

Lecture 7 Artificial neural networks: Supervised learning

Module 2.1 Monitoring activity data for forests using remote sensing

INVESTIGATION LAND USE CHANGES IN MEGACITY ISTANBUL BETWEEN THE YEARS BY USING DIFFERENT TYPES OF SPATIAL DATA

Reification of Boolean Logic

Application of Remote Sensing Techniques for Change Detection in Land Use/ Land Cover of Ratnagiri District, Maharashtra

USE OF RADIOMETRICS IN SOIL SURVEY

A MULTI-RESOLUTION HIERARCHY CLASSIFICATION STUDY COMPARED WITH CONSERVATIVE METHODS

SIMULATION OF FREEZING AND FROZEN SOIL BEHAVIOURS USING A RADIAL BASIS FUNCTION NEURAL NETWORK

DEVELOPMENT OF DIGITAL CARTOGRAPHIC DATABASE FOR MANAGING THE ENVIRONMENT AND NATURAL RESOURCES IN THE REPUBLIC OF SERBIA

A Fractal-ANN approach for quality control

Artificial Neural Networks The Introduction

Choosing Variables with a Genetic Algorithm for Econometric models based on Neural Networks learning and adaptation.

AN INVESTIGATION OF AUTOMATIC CHANGE DETECTION FOR TOPOGRAPHIC MAP UPDATING

Recent Advances in Bayesian Inference Techniques

Calculating Land Values by Using Advanced Statistical Approaches in Pendik

Transcription:

IJCSIT International Journal of Computer Science and Information Technology, Vol. 4, No. 2, December 2011, pp. 55-59 Artificial Neural Network: A Tool for Classification of Land Use and Land Covers Using Satellite Images Manibhushan* 1, Nilanchal Patel 2 and Gadadhar Sahoo 3 1 Scientist (SS), Computer Application in Agriculture, ICAR-Research Complex for Eastern Region, Patna, India 2 Professor, Department of Remote Sensing, Birla Institute of Technology, Mesra, Ranchi, India 3 Professor & HOD, Department of Information Technology, Birla Institute of Technology, Mesra, Ranchi, India ABSTRACT: An artificial neural network is a system based on the operation of biological neural networks, in other words, is an emulation of biological neural system. Artificial Neural Networks or simply Neural Networks are powerful general purpose computing tools. They have become popular in the analysis of remotely sensed data, particularly in classification or feature extraction from image data more accurately than conventional methods [1]. Knowledge of both land use and land cover is very important for socio-economic and agricultural planning of a region. Land use relates to human activities like residential, institutional, commercial, agricultural, recreational, etc. and the land cover relates to the various types of features on the surface of the earth such as river, canal, forest, agricultural land, bare soil, roads, etc. Remotely sensed image data are often used in land cover and land use applications and classification. Classification is a computational procedure that sorts images into groups (classes) according to their similarities. Training has an important role in Neural Network. Land use/ cover classes that include city, water, soil, forest, various types of agricultural lands, buildings, roads, etc. are clearly classified using Artificial Neural Network (ANN). Recent research has therefore focused on obtaining land use / cover information from high resolution satellite and aerial imagery of land use / cover to extract important features from the large amount of information contained in remote sensing data that requires efficient and intelligent analytical techniques. Image classification and feature extraction are critical in classifying land use /cover from either satellite or aerial imagery. Conventional classification methods cannot recognize the phenomena of some spectrum with different land matters so as to degrade classification accuracy. Neural networks do not require a hypothesis about data distribution; they are valuable tools to classify satellite images [2]. 1. INTRODUCTION Land use /cover are an important and highly variable characteristic for effective and efficient future planning. Traditional land use /cover determination and classification are costly, time consuming, tedious and prone to human error. ANN, decision trees, fuzzy logic and neuro-fuzzy techniques have been widely employed recently as an alternative to traditional methods used in classification of large amount of data. Among the more common AI techniques, artificial neural networks (ANN) are computational systems that make use of some of the organizational principles present in biological nervous systems [3]. ANN is a basically pattern recognition device [4] in which the basic computational element is known as the neuron or node. The neuron processes data in to states. First, in combining messages are aggregated by an internal activation function. Then, output from internal activation function is sent to a transfer function, which determines whether or not the neuron will send an output message. Multiple neurons are connected together in layers. These layers are set up to receive input information (input layer), process the data through one or more hidden layers, and produce a corresponding output pattern (through the output layer). ANNs have high processing speed, robustness and generalization capabilities and are able to deal with high dimensional data spaces. The capabilities of ANN for non linear function approximation, data classification, non-parametric regression and non linear decision making are crucial in applications such as multi-spectral image processing, where there is no a priori knowledge regarding data distributions. More particularly, ANN incorporating supervised training algorithms such as feed forward, back propagation networks are capable of distinguishing interesting features from voluminous and noisy data sets having distorted pattern [5]. ANN has been shown to be advantageous over traditional classification methods in achieving higher training accuracy and

56 International Journal of Computer Science and Information Technology generalization ability when dealing with land use /cover related feature classification from a satellite imagery. ANNs have also out-performed simple perceptions or commonly used statistical analysis methods when classifying land use / cover. In order for ANN to be efficient and robust in image classification or feature extraction, pre-processing of data is important. Both spatial and spectral information extracted from images are crucial in using ANNs to distinguish roads from other features in images [6]. A careful organization of data sets to provide crossvalidation tests for ANN training helps to solve the problem of over training [7]. This fact may be extremely important when the available training samples are limited. ANN is also important is obtaining accurate and reliable performance and particularly so when dealing with complicated image processing and classification problems. It is very difficult for a single model or classifier to learn a satisfactory classification rule. This problem may be alleviated by applying an ensemble of classifiers in which a variety of models are learned and combined in an effective way. This technique allows one to use a set of small and simple classifier in a divide and conquer strategy, instead of using one large and complicated model. The simple individual classifiers are combined in an efficient way to yield a lower error rate and better generalization accuracy. The multiple stacking combination scheme [8] is an increasingly studied method of creating so called Committee Machines (CM) from individual ANN expert classifiers. It draws conclusions by simple averaging, weighted averaging, or majority voting from the results of separately trained ANN expert classifiers and is capable of overcoming the problem of over-training, reaching a compromise among conflicting results and increasing generalization ability. A committee machine is a type of neural network using a divide and conquer strategy is which the responses of multiple neural networks (experts) ate combined into a single response. A CM fused knowledge acquired by experts to arrive at an overall decision superior to that attainable by any one classifier s acting alone [5]. The superiority of CMs has been exhibited by their obtaining higher generalization abilities and lower error rates. The combined response of the committee machine is supposed to be superior to those of its constituent experts. 2. METHODOLOGIES USED IN CLASSIFICATION USING ARTIFICIAL NEURAL NETWORK (ANN) Artificial Neural Network or simply Neural Network techniques will give more accurate result in the classification of land use/ cover as compared to traditional methods of classification of satellite images. Some methods of Neural Network techniques for classification of satellite images are being described in the following paragraphs. 2.1. Image Processing and Feature Extraction 2.1.1. Image Acquisition: The first step in image classification is to acquire a digital image. This can be achieved by either using a digital camera for a sensor and digitizer. The sensor device is sensitive to a band in the electromagnetic energy spectrum and produces an electrical signal output in proportion to the energy sensed. The digitizer converts the analogue electrical output into a digital form. To acquire a good image, proper illumination is a basic necessity. The most commonly used image sensors deal with visible and infrared light. 2.1.2. Classification Features: It can be very time consuming if an immense size digital image is to be analyzed in its original form. To make the process of image analysis simple and less time consuming, some qualitative information is extracted from the objects to be analyzed in the image. These extracted objects are called features and a vector of such features is called a pattern. Features are used as inputs to the algorithms for classifying the objects into different categories. Pattern recognition can be done by analyzing the morphology (shape and size), colour, texture (spatial distribution of colours), or a combination of these features of the images. There are two main categories of feature, namely external and internal features [9]. The first step is to extract external features in segmentation, once the objects are separated from the background, their boundary coordinates can be used to extract morphological features, such as Fourier descriptors, boundary chain codes, etc. The features extracted from the properties of pixels inside the object boundary are called internal image features. Depending on the nature of the problem, different types of internal features can be extracted. Spatial moments, colour and textural features are the most important internal image features [10][11].

Artificial Neural Network: A Tool for Classification of Land use and Land Covers using Satellite Images 57 2.1.3. Feature Selection: The success or failure of a classification operation depends on the selection of a feature vector which best describes the given classes. Features that are trivial or are computationally intensive can degrade and slow down the classification results. Optimization set by discarding the redundant and repetitive features is very important to achieve a good classification performance. 2.2. Classifier s Used for Patter n Recognition Classification analysis needs the use of a decision rule, called a classification criterion, to classify objects into two or more known groups, called classes, on the basis of the quantitative features extracted from the objects. A set of features extracted from an object is called an observation of the object. The classification criterion is usually derived from the observation of the known classes, called the training set. The derived classification criterion can then be applied to classify new observations called the test set. Statistical methods (Bayes decision rule) and neural networks are used as pattern classifiers. 2.2.1. Neural Networks as Pattern Classifier: A neural net is a computing network of numerous, simply and highly interconnected processing elements called neurons or nodes. A neuron has many continuous valued input signals x = [xi], I = 1,2,.N, which represent the activity at the input or the momentary frequency of neural impulses delivered by other neurons to this input [12], and an output y which represents the response of the neuron to the input signals. The relationship between the input and output of a neuron is described by the neuron s transfer function, y = f[x]. In the simplest model of a neuron, the output value, y or the frequency of the neuron is often approximated by, 1. : y = f[x] = Kϕ(Σw i x i θ), where K is a constant and ϕ is a non-linear function which takes the value +1 for positive arguments and 1 (or 0) for negative arguments. The wi is called synaptic efficacy [12] or a weight, and θ is a threshold. 2.2.2. Multilayer Neural Networks: A multi-layer neural network (MLNN) with the generalized delta rule for learning by a back propagation learning algorithm [13] is an effective system for learning discriminants for classes from a set of examples [14] [15]. In general, such a network is made up of sets of neurons (nodes) in several layers (Fig. 1). There are three distinct types of layers: the input layer, the hidden layers and the output layer. The connections between the neuron of adjacent layers relay the output signals from one layer to the next. The input layer receives the input information and distributes the information to the next processing layer (the first hidden layer). The number of the neurons in the input layer equals to the dimension of the input vector (the number of features). The hidden and output layers process the incoming signals by amplifying or attenuating or inhibiting the signals through weighting factors. Except for the input layer neurons, the network input to each neuron is the sum of the weighted outputs of the neurons in the previous layer. The number of neurons in the output layer is determined by the number of classes under investigation. The number of hidden layers and the number of neurons in each hidden layer depend on specific application. Figure 1: A Schematic Depiction of a Multilayer Neural Network for N Input and N Output 2.3. Gener al Fr amewor k of the Classification The general framework of the classification procedure includes two parallel classification procedures. The first classifier learned the distribution of land use/ covers types from establishing spatial relationships between the land use / cover (outdated) map and ancillary data [16]. This allows the production of digital fuzzy map which portrayed, via each pixel, the likelihood of the presence of each land use/ cover type. The second classifier produces another fuzzy map using a spectral classification of the recent remotely sensed image. Thus, for each pixel, the two fuzzy maps indicate a membership value which expresses the likelihood of the presence of cover type from its environmental conditions and its spectral features, respectively. In order to combine the two fuzzy maps, the AND operator is used to calculate the minimum of two membership values. The use of the AND operator ensures that the most stringent requirement for the class

C o n t o u r s a n d W a v e f o r m s. IEEE 58 International Journal of Computer Science and Information Technology section is met. A final hard (no fuzzy) map is finally obtained labeling each pixel into the class with the higher fuzzy membership value. In order to carry out, the classification based up on the ancillary data, each spatial variable is overlaid on the land use/ cover map from INEGI (National Institute of Geography, Statistics and Informatics) to establish the relationship between the land use and cover and the variables. These data are used to train the first MLP (Multilayer Perception) aimed at classifying land use/ cover from the environmental variables. The spectral classification is carried out following two steps. First, an unsupervised spectral classification is carried out using the ISODATA algorithm. This classification allows the selection of training sites which represents the whole spectral variety a class could represent and are used to train the second MLP. A back propagation algorithm is used in the MLP training processes. Data are divided into three sections: the training set, the verification set and the test set. Training algorithms do not use the verification or test sets to adjust networks weights. The verification set is used to track the network s error performance, to identify the best network, and to stop training if over-learned occurred. The test is not used in training at all, and gives an independent assessment of the network s performance when the entire network design procedure is completed. The network configuration is determined empirically by testing various possibilities and evaluating the accuracy of the classification of the test set. Among the MLP architectures which present good performance, the simplest are chosen (less input variables and fewer nodes in the hidden layer) because of their better ability to generalize and classify unseen pixels accurately. In order to assess the accuracy of the classified images, a random sample of 488 points of verification is selected. The land use/ cover classes of surrounding area of these points are checked by visual interpretation using high resolution digital aerial photographs. A matrix of error is constructed and accuracy indices are computed [16]. In addition, it is expected that the range of applications of neural networks in remote sensing will broaden. Applications in which neural networks have already been used and increased usage may be expected include image processing (e.g. geometric, atmospheric and radiometric correction, stereomatching imagery, image compression, feature extraction, map generalization, multi-source data analysis, data fusion and image sharpening [17]. The computing world has a lot of gain from neural networks. Their ability to learn by example makes them very flexible and powerful [18]. 3. CONCLUSION Artificial Neural networks are powerful computing tools. They have been used in the classification of remotely sensed image data and regression type problems in which they have often been demonstrated to extract information more accurately than conventional methods. Although not free from problems it seems likely that the neural networks will be used increasingly in ecological research using remote sensing. Moreover, as some of the problems encountered in the use of neural networks arise from a tendency to focus upon the MLP only it is likely that there will be a greater use of other network types. A large number of claims have been made about the modeling capabilities of neural networks. Thus while neural networks have rapidly become established in remote sensing it is likely that they will be used increasingly and a broader range of activities that will help to exploit more fully the potential of remote sensing as a useful tool in ecological research. REFERENCES [1] G. Recknagel Friedrich and Foody M., (2006), Pattern Recognition and Classification of Remote Sensed Images by Artificial Neural Networks, Ecological Informatics Springer Berlin, Heidenberg, May, 2006. [2] Monica Bocco, Gustava Ovando, Silvina Ayago and Enrique Willington, (2007), Neural Networks Models for Land Cover Classification from Satellite Images, Agriculture Technica (Chile) 67(4), 414-421, Cordoba, Argentina. (Oct-Dec, 2007). [3] Du Bose and Klimasaukas (1989), Introduction to Neural Networks with Examples and Applications, Neuralware Inc, Pittsbergh. [4] Sui and Thomasson (2006), Ground based Sensing System for Cotton Nitrogen Status Determination, Trans ASABE. [5] Haykin, (1999), Neural Computing, Second Edition, PHI, Princeton, NJ. [6] Boggess, (1994), Identification of Roads in Satellite Imagery using Artificial Neural Networks: a Contextual Approach In: Proc. World Congress Neural Networks, INNS, Madison, WI, pp. 1410-1415. [7] Mitchell, (1997), Machine learning, Mc.Graw-Hill, Network, NJ. [8] Wolpert David H. (1992), Stacked Generalization, Neural Networks, 5(2), 241-259. [9] Pavlidis T., (1980), Algorithms for Shape Analysis of Transactions on Pattern Analysis and Machine Intelligence. 2(4), 301-312. [10] Gonzalez R. C, Woods R. E., (1992), Digital Image Processing, Addison Wesley Publishing Co., Reading MA. [11] Levine M. D. (1985), Vision in Man and Machine. Mc Graw Hill Inc., New York, NY.

Artificial Neural Network: A Tool for Classification of Land use and Land Covers using Satellite Images 59 [12] Kohonen T. (1988), An Introduction to Neural Computing, Neural Networks, 1, 3-16. [13] Rumelhart D. E., Hinton G. E., Williams R. J. (1986), Learning Internal Representation by Error Propagation. In: Parallel Distributed Processing, 1, 318-362 MIT Press Cambridge MA. [14] Sejnowski T. J., Rosenberg C. (1987), Parallel Networks that Learn to Pronounce English Text, Complex Systems, 1,145-168. [15] Tesauro G., Sejnowski T. J. (1989), A Parallel Network thatlearn to Play Backgammon. Artificial Intelligence 39, 357-390. [16] Mas Jean Francois, (2003), An Artificial Neural Networks Approach to Map land use/ cover using LANDSAT Imagery and Ancillary Data, Instt. of Geografia, IEEE 2003. [17] G. M. Foody (1999), Department of Geography, University of Southampton, Highfield, Southampton, SO 17 1B.J. U. K. [18] Jha. Girish Kumar., Artificial Neural Networks and its applications, IARI, New Delhi.