ANN based GUI to Classify Satellite Images for Remote Sensing

ANN based GUI to Classify Satellite Images for Remote Sensing

Abstract: Automatic classification of satellite images to detect the presence of urban/forested/deserted features of the areas on earth plays a most crucial part in the field of remote sensing. This paper presents an image classification technique to classify the urban/forested/deserted areas on input satellite images by utilizing mean-shift clustering with artificial neural network. The input satellite images are first enhanced through fuzzy histogram equalization and the Haralick's texture features are utilized to ascertain the arbitrariness in the image. Testing and training is done through the ANN and the resultant image is finally utilized to quantify its accuracy and error rate. The proposed algorithm is implemented on MATLAB GUI and obtains a classification accuracy of 90.9% and is immune to noise.

Keywords: Artificial Neural Network, Image Classification, MATLAB, Mean Shift Clustering, Remote Sensing

INTRODUCTION

A digital image is the representation of a two dimensional image as a finite set of digital value called picture elements or pixels. On the basis of number of pixels, it may be classified as High Resolution, Medium Resolution or Low Resolution Image [1]. Today, Digital Image Processing is being used for research & development purpose in almost all the important areas. One of such area is Remote Sensing. Use of image classification is important in remote sensing to produce land cover maps and to suggest the information about the usage of land in terms of Forests/Urban Land/Desert/Ocean etc [2]. As this information is to be classified from the images that are to be acquired from satellites, hence, it is mandatory to develop an algorithm which accurately classifies such information from high-resolution satellite image. A sample high-resolution satellite image is shown in figure 1 below:

Figure 1: High Resolution Satellite Image [3]

It is also pertinent to note that Artificial Neural Network (ANN) is widely used for feature classification in digital images. The prime advantage of using ANN is its capability to handle high resolution images with less computational complexity and best error tolerance.

Literature Review

Shohel Ali Ahmed et al. [4] presented an ANN based texture classification or segmentation which is an advanced technique to provide rich information of an image. It works suitably on remotely sensed images and can extract objects from images that are without complex backgrounds i.e. dont have major colour or textural variations.

D. Chaudhuri et al. [5] presented an approach to extract roads from high resolution panchromatic remotely sensed imagery. Morphological segmentation & template matching is used to detect only the presence of roads in satellite images. This algorithm requires at least 13 stages

including road enhancement, road segmentation, hole filling, small region filtering, length based region filtering, small branch removal method and road segment linking [5.

1. Ankayarkanni et al. [6] suggested a way to deal with the extraction of road and structures from a high resolution satellite image and compared the performance of k-means clustering algorithm to the KFCM strategy. It is finally suggested by the author that KFCM have better performance over K-means clustering method.

   Capt. Dr.S Santosh Baboo et al. [7] presented an approach for effectively detecting the features on the surface of Earth i.e. the places that are recognized from the scalable imagery using FCM (Fuzzy C-means). This method generates the segmented results of selected regions in image but the use of FCM is rare & ineffective for practical usage due to complex spatial correlation properties of image pixels.

   S K Katiyar et al. [8] proposed a neural network based approach to extract and identify various objects like road network, water body & lake from satellite images. A model of back propagation neural network BPNN combined with Haralicks texture feature is used to classify these objects.

   Rongjun Qin [9], proposed a mean shift vector based shape feature (MSVFS), a noval spatial feature which enhances the classification accuracy of very high resolution remote sensing imagery Support vector machine (SVM) in used to group spectral and spatial elements.

   It is observed from the literature review that existing ANN based segmentation system is incapable of extracting objects from complex backgrounds when the colour and textural variations are observed [10]. It is further observed that although FCM is widely used technique for satellite image segmentation but its effective practical use is rare due to spatial correlation properties of image pixels [11]. In addition to this, spatial segmentation techniques (such as watershed algorithm) based segmentation undesirably produces a large number of small quasi-homogenous regions and finally, edge information is not preserved in case of clustering based segmentation approach. This paper presents an approach to develop an algorithm by combining the advantages of MST based clustering with ANN to overcome these drawbacks.

   Research Methodology

   Figure 2 below shows the flowchart of the proposed algorithm. The algorithm is tested over 44 high resolution satellite images that are principally acquired from online database and Google Earth. The images used for testing are the blend of rural, urban and deserted territories.

   Fuzzy Histogram Equalization

   The input images are further enhanced for their contrast using Fuzzy Histogram Equalization (FHE) [12]. The FHE comprises of two phases. To start with, fuzzy histogram is figured taking into account fuzzy set hypothesis to handle the estimation of grey level qualities better contrasted with established crisp histograms. Secondly, the fuzzy histogram is isolated into two sub-histograms in light of the median estimation of the first image and after that evens out them freely to protect image brightness. The subjective and quantitative investigations of proposed FHE calculation are assessed utilizing two parameters, average information content (AIC) and natural image quality evaluator (NIQE) file for different pictures. Experimental result demonstrates that the proposed technique can successfully and fundamentally dispense with washed-out appearance and antagonistic curios instigated by a few existing strategies. Mathematically, it is represented as given in equation 1.

   = =1 =1 [0,1]

   Eq. (1)

   Where, gmn is the intensity value of (m,n)th pixel and µmn is its membership function.

   Input satellite image

   Fuzzy histogram equalization

   Fuzzy histogram equalization

   Mean shift clustering

   Mean shift clustering

   Determining

   Haralicks feature

   Determining

   Haralicks feature

   Accuracy error rate and noise sensitivity

   Accuracy error rate and noise sensitivity

   Image classification

   Image classification

   Training and testing through artificial neural network

   Training and testing through artificial neural network

   Figure 2: Flowchart of proposed algorithm

   Mean Shift Clustering

   Mean Shift clustering acts as a mode finding algorithm [13]. Every one of the focuss in the same bowl of interest is connected with the same group. The quantity of cluster is obtained by the quantity of modes. It is capable to hold the notable elements of the general images because of its edge saving sifting property.

   Haralicks Features

   It is a statistical measure that defines the correlation between pixels that are presented in neighbourhood in a given space [14]. Out of the fourteen features that determine the textural or surface properties of an image, entropy is the one. The more the entropy, the more likely it is to be a complex image. It measures the randomness in the image and is defined as,

   Here, the parameter q is the likelihood that the distinction between two neighbouring pixel is equivalent to (m,n) and is the logarithmic capacity.

   = , (, ) log((, ))

   RESULTS & DISCUSSION

   Eq. (2)

   Artificial Neural Network

   The Artificial Neural Networks are generally unrefined electronic models in view of the neural structure of the mind [15]. In this paper, a back propagation neural network (BPNN) is employed with nine information layers and ten hidden layers to classify the picture by grouping them as pixels belonging to rural or urban or some deserted image.

   The proposed algorithm has been designed & tested over MATLAB. A GUI is designed for effective representation of results. The testing has been done over 22 satellite images under two scenarios: Noiseless and Noisy. Overall the results have been presented in this paper for 44 images. Figure 3 below shows four different types of sample images taken for testing, and their corresponding results obtained after image enhancement and classification using ANN are also presented.

   1. (e) (i)

   2. (f) (j)

   3. (g) (k)

   4. (h) (l)

Figure 3: (a)-(d) shows the input satellite images, (e)-(h) shows the results of image enhancement using FHE and (i)-(j) shows the results of image classification on MATLAB GUI

Table 1 below presents the expected & obtained results of image classification using proposed algorithm in noise-less images and Table 2 summarises the similar results for same input images when corrupted by noise.

Table 1: Image Classification Results obtained on Noise-less Satellite Images

Table 2: Image Classification Results obtained on Noisy Satellite Images

Table 3: Scenario Wise Quantitative Analysis of Results

Comparison of results

Table 4 below represents the accuracy result of the proposed algorithm with the existing algorithms and it is been found that the accuracy of the proposed algorithm is more as compare to the existing algorithm [16].

Table 4: Accuracy (in % ) comparison representation

CONCLUSION

The work gives an edge work of utilizing ANN for picture grouping and division. The framework as executed in mat lab programming is computationally simple and fast. Results of the algorithm are found to be more accurate than the methods like minimum spanning tree or watershed. The framework distinguishes the components accurately as it uses the Haralicks textures, state of the items withstanding pixel values, colour and texture feature for the classification of result. Which is classifying the information as a urban, rural or neither rural nor urban images. The future upgrades of the framework incorporate improvement of the framework by the utilization of setting data and general standards for image investigation.

REFERENCES

1. Karl Segl and Hermann Kaufmann, Detection of small objects from high resolution panchromatic satellite imagery based on supervised image segmentation, IEEE Transactions on Geoscience and remote sensing, 39(9), page no2080-2083, 2001.

2. Aaron k Shakelfold , A combined fuzzy pixel based and object based approach for classification Of high resolution multispectral data over urban areas, IEEE transactions on geoscience and Remote sensing, vol 41,no. 10,2003.

3. T Rajni Mangla and S G Birud, A New Auotomatic Road Extraction technique using gradient Operators and skeletal ray formation, International Journal of Computer Applications (0975 8887)

   Volume 29 No.1, September 2011

4. Shohel Ali Ahmed Snigdha Dey and Kandarpa Kumar Sarma, Image texture classification using Artificial neural network, IEEE 2nd National Conference on Emerging Trends and Applications In Computer Science (NCETACS),4-5 March 2011, Shillong, ISBN:978-1-4244-9578-8,pp:1-4.

5. D.chaudhuri ,N.K. Kushwaha and A.samal, Semi-automated road detection from high resolution Satellite images by directional morphology enhancement and segmentation techniques, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2012, pp: 1538- 1544

6. B.Ankayarkanni , Ezil Sam Leni, Object based segmentation technique for classification of satellite image, Indian Journal of Computer Science and Engineering (IJCSE), ISSN : 0976-5166 Vol. 5 No.3 Jun-Jul 2014, pp:119-125

7. Capt. Dr. SS Baboo, Mr. Sthirunavukkarasu, Image Segmentation using High Resolution Multispectral Satellite Imagery implemented by FCM Clustering Techniques, IJCSI International Journal of Computer Science Issues, Vol. 11, Issue 3, No 1, May 2014, ISSN (Print): 1694-0814 | ISSN (Online): 1694-0784, pp:154-161. S K Katiyar and P V Arun, An enhanced neural network based approach towards object Extraction" ,IEEE Transactions on geosciences and remote sensing, vol. 50 No.11b

8. Rongjun Qin, Student Member, A Mean Shift Vector-Based Shape Feature for Classification Of High Spatial Resolution Remotely Sensed Imagery ,IEEE Journal of selected topics in Applied earth observation and remote sensing (J-STARS), ISSN: 1939- 1404, September 2014,pp:1-12.

9. P. SathyaandL.Malathi, Classification and Segmentation in Satellite Imagery Using Back Propagation Algorithm of ANN and K- Means , International Journal of Machine Learning And Computing, vol. 1, No. 4, October 2011, pp: 422-426.

10. Ujjwal Malik and Sanghamitra Bandyopaddhaya, fuzzy partitioning using a real coded variable Length genetic algorithm for pixel classification,IEEE transactions on geosciences and remote Sensing, vol 41, no 5,may 2003,page no 1075-1079.

11. Mahendra PS kuber, Manish Dixit and Sanjay Silakari, Improving brightness using dynamic Fuzzy histogram equalization international journal of signal processing, image processing and Pattern recognition vol 8 No (2015) pp:303-312

12. P. Padmapriya and K. Thenmozi , Agglomerative mean shift cluster using shortest path and fuzzification algorithm, international journal of advanced research in computer science and software engineering ISSN:2277 128X vol 3,Issue 12 ,December 2013 ,page no 919-930.

13. Gebejes and R,Huratas , Textural classification based on grey level co- occurrence matrix Conference of informatics and management sciences march 25,-29 2013.

14. [15] Gowri Ariputhiran1, S. Gandhimathi Usha, Feature Extraction and Classification of High Resolution Satellite Images using GLCM and Back Propagation Technique", International Journal of Engineering and Computer Science ISSN: 2319-7242 Volume 2 Issue 2 Feb 2013 page no 525-528

15. Biplab Banrjee, Surender verma, Krishna mohan and Laxmi Narayan Eeti, Unsupervised multi Spectral satellite image segmentation combining modified mean shift and a new minimum Spanning tree based clustering technique, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (J- STARS), ISSN:1939-1404, March 2014, pp:888 894