Performance Analysis of Feature Extraction Techniques Applied on SAR Data

Performance Analysis of Feature Extraction Techniques Applied on SAR Data

L.Subathra

Dept. of Electronics & Communication Engineering Thiagarajar College of Engineering

Madurai, India

R. A. Alagu Raja

Dept. of Electronics & Communication Engineering Thiagarajar College of Engineering

Madurai, India

AbstractFeature extraction in image processing is used for extracting certain features of interest such as edge, shape and texture features. These features are necessary to classify the image into major classes like water body, urban, vegetation and others. This analysis is done on Synthetic Aperture Radar (SAR) image of Madurai region. The image is acquired from Sentinel- 1A satellite. In this paper, the performance of two feature extraction techniques are analyzed with respect to classification. Image analysis is limited by the inherent noise present in SAR images, known as speckle. Preprocessing is necessary to remove the speckle noise. For this purpose, Lee filter is applied on the image. For feature extraction, many spectral based methods are used. Among them, Laws mask and Gabor filter based feature extraction techniques are discussed in this paper. Laws mask is a bank of filters, that is appropriate for texture identification. Among the 25 masks, Level mask, Spot mask, Ripple mask, Edge mask and Wave mask are implemented here. Gabor filter is a linear filter used to represent and to discriminate texture features. The extracted features are grouped into different classes using classification techniques. The k-means and fuzzy c means (FCM) clustering are the two unsupervised classification techniques discussed here. The Gabor filter provides better feature information compared to Laws mask. FCM is preferred over k-means clustering since it provides good results with both Laws mask and Gabor Features. The experimental result shows that the Gabor filter with FCM provides high accuracy of 74 %.

Keywords Feature Extraction, Classification, Laws mask, Gabor filter, k-means, Fuzzy c means

I. INTRODUCTION

Feature Extraction is a method used in image processing as a segmentation step [19]. Feature extraction is done after pre- processing [11]. In this process, the image is converted into features which give information about the image [4]. Features contain some information to distinguish between classes [1]. The features may be a spatial feature, a histogram feature, a transform feature, an edge and boundary feature, a shape feature, a color feature or a texture feature. Using Laws mask and Gabor filter, texture features are extracted [12]. The features extracted from different methods are then converted into feature vectors and then used for further processing [7]. In this paper, the feature vectors are then classified so that the image can be classified into four major classes like water body, Urban, Vegetation and Others. It is easier for the classifier to classify the image using the features since classes can be easily distinguished from feature extraction step [15]. In this paper, the main focus is on feature extraction. But classification is essential to classify the land into four major

classes. Many approaches have been used till now to extract features. Here, two spectral based methods are discussed and the performance of these methods is analyzed and the results are shown in the upcoming sections.

1. Study area and Dataset

   A Synthetic Aperture Radar (SAR) is a coherent side- looking radar system which generates high-resolution remote- sensing imagery. Synthetic aperture radar (SAR) is a form of radar that is used to create two or three dimensional images of objects such as landscapes. SAR uses the motion of the radar antenna over a target region to provide finer spatial resolution than conventional beam-scanning radars [6]. SAR images have wide applications in remote sensing and mapping of the surfaces of both the Earth and other planets [3]. Good range resolution relies principally upon the properties of the transmitted waveform. SAR operates in four modes: Strip Map mode, Interferometric Wide Swath mode, Extra-wide swath mode and Wave mode [17].

   The implementation is done on Madurai city, India. For undergoing the survey, the range covered over Madurai region is 8 km. The SAR image of Madurai is obtained from the SENTINEL-1A, a European radar imaging satellite. Sentinel-1 is a two satellite constellation with the prime objectives of Land and Ocean monitoring. The satellite carries a C-band Synthetic Aperture Radar which will provide images in all light and weather conditions. Sentinel-1 operates in C band. The satellite frequency is 12-18 GHz (ku band). The Sentinel-1A is launched on April 3, 2014. The instrument used in this satellite is SAR-C. Sentinel-1 satellite is about 693 km above the earth surface. It covers up to 400 km. It is a Sun-synchronous satellite. The incident angle is 20°-45°. Transmitting frequency is 5.8 to 6.4 GHz and the receiving frequency is 3.6 to 4.2 GHz. The mode used for capturing this image is Interferometric Wide Swath (IW) mode. The swath width is 250 km. The polarization is VV and VH.

   Some of the applications for launching Sentinel-1A satellite are (a) Monitoring sea ice zones and the Arctic environment, and surveillance of marine environment, (b) Monitoring land surface motion risks, (c) Mapping of land surfaces: forest, water and soil and (d) Mapping in support of humanitarian aid in crisis situations.

   Fig.1 depicts the SAR image of Madurai, taken on July 2016 with the spatial resolution of 20*22 m.

   Fig.1. SAR Image of Madurai, India

   Lee filter calculation produces an output value close to the original input value in higher contrast regions and a value close to the local mean for uniform areas [4]. In uniform areas more smoothening occurs. The Lee filter follows the below equation,

   Img (i,j) = Im + W * (Cp – Im) (1)

   where, Img is the pixel value after filtering

   Im is the mean intensity of filter window Cp is the center pixel

   W is the filter window

   In the next section (Section II), the methodology of the proposed work was illustrated. It explains in detail, the

   following: 1) the preprocessing of the data before being used for evaluation; 2) the feature extraction methods to be applied, such as Laws mask and Gabor filter, and, finally, 3) the classification methodology to be used for the evaluation of two cases. In Section III, experimental results using the Sentinel-1A SAR image of the Madurai site, are demonstrated. Finally, some concluding remarks are given in Section IV.

   II PROPOSED METHODOLOGY

   The general approach outlines here is to extract the features from an image so that it is easy for classification. Fig.2. shows the basic block diagram of the proposed work. It involves three main steps: Preprocessing, Feature Extraction and Classification.

   Fig.2 Methodology of the proposed work

   Every block of the proposed work is necessary to be implemented to obtain the classified output.

   1. Preprocessing

      Pre-processing is essential to improve the Visual Interpretability [9]. In this step, the image is enhanced by removing noise. Radar image appears more noisy than an optical image. SAR image is highly affected by speckle noise [8]. Speckle noise occurs due to the coherent summation of the signals scattered from ground scatterers distributed randomly within each pixel. Speckle is a granular noise that inherently exists in and degrades the quality of the active radar, synthetic aperture radar [20], medical ultrasound and optical coherence tomography images. In order to reduce specklein SAR images, some adaptive speckle filters namely Gamma, Lee, frost, Enhanced Lee (EL) and Enhanced Frost (EF) filters are used. Among these,

      1. (b)

         Fig.3. (a) Original Image (b) Preprocessed Image

         The Fig.3 depicts the speckle noise removal using Lee filter. The preprocessed image contains more information compared to the input image in Fig. 1.

   2. Feature Extraction

      Feature Extraction can be carried out using Statistical, Structural and Spectral based methods. Among them, Spectral-based feature extraction methods are applied in the proposed work. Since spectral methods analyze the frequency content of an image, it gives more spectral information of an image [16]. The two methods such as Laws mask and Gabor filters are used.

      Laws mask

      Laws mask is an approach for texture identification using filtering. Derived from the concept of texture energy defined at each pixel after a series of particular convolution with selected mask, laws mask can produce the texture energy measurement for the analysis of the texture property of each pixel [12]. Among 25 filters, the most commonly used masks are Level mask, Edge mask, Spot mask, Wave mask and Ripple mask.

      Level mask identifies any variations in a homogeneous region of the image to which it is convolved with, whereas convolving with edge mask, the edges or abrupt changes in the input image can be extracted. Convolution of spot mask identifies any point level changes in the input image. Ripple mask extracts any surface texture roughness in the given imagery. And the convolution with wave mask extracts any periodic variations in the image.

      Fig.4 demonstrates the features extracted using Laws mask. Each feature gives unique information. While classifying the image, each feature has to be converted to feature vector.

      Gabor filter can achieve the best localization performance both in time and frequency domain through analyzing the uncertainty of signal in direction and spatial frequency [10].

      Gabor Filter

      Fig.4. Features extracted using Laws mask

      A Gabor function is a sinusoidal signal with a given frequency and orientation, modulated by a Gaussian function [5]. The filter used here is called Gabor filter. The Gabor filter bank is defined by its parameters including frequencies, orientations and smoothing parameters of the Gaussian envelope [13]. The Gabor filter is obtained by using different scales and orientations. A complex 2-D Gabor filter over the image domain (x, y) is defined as,

      (2)

      where, is the spatial spread

      =2f, f is the radial frequency is the orientation

      If a set of Gabor filters with different orientations and frequencies are used, then a host of features can be extracted [14]. Totally, 32 filters are obtained for four scales and eight orientations.

      The features extracted using Gabor filters for all the four scales and eight orientations are depicted in figures 5,6,7 and 8.

      Fig.5. Features extracted for scale f=0.05

      Fig.6. Features extracted for scale f=0.1

      Fig.7. Features extracted for scale f=0.2

      Fig.8. Features extracted for scale f=0.4

   3. CLASSIFICATION

Image classification is a significant tool for digital image analysis and object recognition. The major steps involved in image classification are determination of suitable classification system, selection of training and testing samples and the classification technique [2].

k-means Clustering

k-means algorithm is a well known clustering algorithm popularly known as Hard c means algorithm. It is an algorithm to classify or to group the objects based on attributes/features into k number of group. k is positive integer number [7]. The grouping is done by minimizing the sum of squares of distances between data and the corresponding cluster centroid.

1. eans clustering aims to partition n observations into k clusters in which each observation belongs to the cluster with the nearest mean, serving as a prototype of the cluster [18]. This algorithm splits the given image into different clusters of pixels in the feature space, each of them defined by its center. Initially each pixel in the image is allocated to the nearest cluster. Then the new centers are computed with the new clusters.

   Fuzzy c means Clustering (FCM)

   Fuzzy Clustering is a soft clustering. In Soft Clustering, data elements can belong to more than one cluster and a membership level associated with each element indicates the strength of that element with a particular cluster [9]. In this algorithm, membership levels are assigned to each pixel and

   then it is used to assign data elements to clusters.

   Main objective of fuzzy c-means algorithm is to minimize,

   (3)

   Let X = {x1, x2, x3 …, xn} be the set of data points and V =

   {v1, v2, v3 …, vc} be the set of centers.

   • Randomly select c cluster centers.

   • Calculate the fuzzy membership 'µij' using:

     (4)

   • Compute the fuzzy centers 'vj' using:

     (5)

   • Repeat the steps until the minimum 'J' value is achieved or .

III EXPERIMENTAL RESULTS

Classification approaches can be implemented to distinguish one or more specific classes of terrain (such as water bodies, paved surfaces, irrigated agriculture, forest cutting, or other types of disturbances) within the landscape. The two unsupervised classification techniques used in this paper are k-means and fuzzy c means (FCM). In the classification, the land cover is classified into four classes as urban, water body, vegetation and others represented by red, blue, green and yellow respectively. Fig. 9 depicts the classification result of laws mask feature.

1. (b)

Fig.9. k-means Classification for (a) laws mask feature. (b) Gabor feature

Classification approaches can be implemented to classify the total scene content into a limited number of major classes [15]. Fig.10 depicts the classification result of Gabor feature.

(a) (b)

Fig.10. FCM Classification for (a) laws mask feature (b) for Gabor feature

Accuracy Assessment

A Classification is not complete until its accuracy is assessed [6]. Construction of confusion matrix or error matrix is a best way to assess the accuracy of a classification. An accuracy table usually provides the performance result of a classification in terms of overall accuracy, producer accuracy and user accuracy. The performance evaluation is done by calculating the accuracy using the confusion matrix and is tabulated. The classified images are assessed for the accuracy by comparing with Google Earth ground truth. Optimal classified outputs are only taken for the accuracy assessment by computing confusion matrix. Totally, 100 sample points are taken and the accuracy is calculated.

W-Water body, U-Urban, V-Vegetation, X-User accuracy, Y-Producer Accuracy.

TABLE I LAWS MASK AND K-MEANS

Overall Accuracy=56% TABLE II LAWS MASK AN FCM

Overall Accuracy=59%

TABLE III GABOR FILTER AND K-MEANS

Overall Accuracy=71%

TABLE IV GABOR FILTER AND FCM

Overall Accuracy=74%

User accuracy is a measure of the reliability of an output map generated from a classification scheme. Producer accuracy is a measure of the accuracy of a particular classification scheme. Overall accuracy is the percentage of correctly classified pixels.

Diagonals represent sites classified correctly according to reference data and Off-diagonals represent misclassified

pixels. Overall accuracy is the percentage of correctly classified pixels.

IV CONCLUSION

Classification is done on the basis of extracted feature vectors in to four classes namely water bodies, vegetation, urban and others. The classified images are assessed for the accuracy by comparing with Google Earth ground truth. Optimal classified outputs are only taken for the accuracy assessment by computing confusion matrix. The feature vectors obtained using the two techniques are fed independently to the two classification mechanisms. Thus, four different results are obtained. On analyzing the results using the accuracy assessment table, the following conclusions are made. The Gabor filter provides better feature information compared to the Laws Mask. FCM has robust characteristics for ambiguity and it can retain more information than k-means clustering [14]. It gives best result for overlapped data set and comparatively better then k- means algorithm. Hence, Gabor filter with FCM produces higher accuracy (74%) compared to other methods.

REFERENCES

1. Hongying Liu, Shuyuan Yang, Shuiping Gou, Dexiang Zhu, Rongfang Wang, and Licheng Jiao, Polarimetric SAR Feature Extraction With Neighborhood Preservation-Based Deep Learning, IEEE Journal Of Selected Topics In Applied Earth Observations And Remote Sensing, October 2016.

2. Jiawei Chen, Licheng Jiao and Zaidao Wen, High-Level Feature Selection With Dictionary Learning for Unsupervised SAR Imagery Terrain Classication, IEEE Journal Of Selected Topics In Applied Earth Observations And Remote Sensing, February 2016.

3. Liang Cheng, Yafei Wang, Lishan Zhong,Peijun Du and Manchun Li , Technical Framework of Feature Extraction based on Pixel- Level SAR Image Time Series, IEEE Journal of Selected Topics in Applied Earth Observation and Remote Sensing, vol.8, no.4, April 2015.

4. F. Albregtsen, Statistical texture measures computed from gray level co-ocurrence matrices, University of Oslo, Oslo, Norway, 2008, Image Processing Laboratory.

5. Song-lin Liu, Zhao-dong Niu, Gang Sun, Zeng-ping Chen, Gabor Filter-based Edge Detection: A note, ELSEVIER, 2014.

6. Corneliu Octavian Dumitru and Mihai Datcu, Information Content of Very High Resolution SAR Images: Study of Feature Extraction and Imaging Parameters IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 51, NO. 8, AUGUST 2013.

7. C. O. Dumitru and M. Datcu, Study and assessment of selected primitive features behavior for SAR image description, in Proc. IEEE IGARSS, Munich, Germany, 2012, pp. 35963599.

8. Christina Corbane, Nicolas Baghdadi, Xavier Deccombes, geraldo Wilson Junior, Nicolas Villeneuve, and Michel Petit, Comparative Study on the Performance of Multiparameter SAR Data for Operational Urban Areas Extraction using Textural Features, IEEE Geoscience and Remote Sensing Letters, vol.6, n0. 4, October 2009.

9. N. Suresh Kumar and M. Arun, Enhanced Classification Algorithms for the Satellite Image Processing, Indian Journal of Science and Technology, Vol 8(15), July 2015.

10. Du-Ming Tsai, Song-Kuaw Wu, Mu-Chen Chen, Optimal Gabor filter design for texture segmentation using stochastic Optimization, ELSEVIER 2000.

11. Usman Babawuro, Zou Beiji, Satellite Imagery Cadastral Features Extractions using Image Processing Algorithms: A Viable Option for Cadastral Science , IJCSI International Journal of Computer Science Issues, Vol. 9, Issue 4, No 2, July 2012.

12. Heba A. Elnemr, Statistical Analysis of Laws Mask Texture Features for Cancer and Water Lung Detection , IJCSI International Journal of Computer Science Issues, Vol. 10, Issue 6, No 2, November 2013.

13. Ma Li, R.C. Staunton, Optimum Gabor lter design and local binary patterns for texture segmentation, ELSEVIER 2007.

14. Rajni Rani, Kamaljeet Kaur, Implementation for Gabor Filter Using on Satellite Images Enhance the Image Quality, International Journal of Advanced Research in Computer Science and Software Engineering, Volume 3, Issue 5, May 2013.

15. Lance M. Kaplan, Extended Fractal Analysis for Texture Classication and Segmentation, IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 8, NO. 11, NOVEMBER 1999.

16. Xiangrong Zhang, Licheng Jiao, FangLiu, Liefeng Bo, and Maoguo Gong, Spectral Clustering Ensemble Applied to SAR Image Segmentation, IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 46, NO. 7, JULY 2008.

17. Mattia Stasolla, and Paolo Gamba, Spatial Indexes for the Extraction of Formal and Inormal Human Settlements From High- Resolution SAR Images, IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 1, NO. 2, JUNE 2008.

18. F. DellAcqua, P. Gamba, G. Trianni, Semi-automatic choice of scale-dependent features for satellite SAR image classication, ELSEVIER, 2005.

19. Mausumi Acharyya, Rajat K. De, and Malay K. Kundu, Segmentation of Remotely Sensed Images Using Wavelet Features and Their Evaluation in Soft Computing Framework, IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 41, NO. 12, DECEMBER 2003.

20. B. Liu, C. Zhu, K. Wang, X. Liu, and W. Yu, A one-class- extraction framework for high resolution SAR image classication, in Proc. IEEECIE Int. Conf., 2011, pp. 732735.