Satellite Image Classification based on Ensemble Subspace Discriminant method using Random Subspace Algorithm

Author Name(s): K.Radhika, S.Varadarajan
Author Email: krkvarma15@gmail.com

Abstract

Classification of images obtained from satellites is a challenge now-a-days in order to know the information contained by them. This requires the adaption of the present state-of-art classifiers. Correct information can be retrieved only with good classification method. But single classifier may not meet the requirement, so that ensemble techniques are preferable. For classification with three or more classes various methods are AdaboostM2, LPBoost, TotalBoost,RUSBoost,Subspace,Discriminantetc etc. It is possible to create an ensemble for classification by combining any of these. Out of these, we have considered random subspace (RS) and discriminant classifiers for ensemble. Here bagging method preferred. This ensemble can be used for weak learners also. To judge whether the classifier is giving correct information or not validation is important. To validate this ensemble technique we have considered 50,100,150 and 200 sample points, examined the accuracy and necessary parameters. The nature of the proposed method for the above said number of sample points is discussed.

Introduction

Classification is playing major role in the extraction of information contained by the satellite images. During this processof classification several problems may come across like accuracy, misclassification etc. There are several classifiers like Support vector machine,K nearest neighbors, neural network, maximum-likelihood  [1-3].But to improve accuracy and to reduce misclassification single classifier is not sufficient. To strengthen the classification efficiency two or more base classifiers need to combine.This is possible by ensemble of several classifiers [4-6].

The prediction capability of the ensemble classifier model can be increased by the attributes. These will support training processof classifier in case of uncertain tuples thereby correct prediction possible. With all these procedure, a high accuracy is possible by ensemble classifiers compared to single classifiers. Ensemble procedure consists of bagging, boosting. Bagging process will generate various training data. With repeated sampling a new training sets can be derived from original sets. If we repeated N times, N number of new data sets will be generated.After this, boosting is required to enhance the quality of the ensemble model[7]. Random subspace is the majorly using ensemble strategy and this procedure is taken place in feature space. To get final data set majority voting is preferred. For boosting the algorithm used generally is AdaBoost. This will combine a set of weak classifiers into a strong classifier. Weak classifiers are nothing but whose accuracy is lesser than expected. Boosting algorithm will modify the weight instances to make the classifiers stronger.The high accurate classifiers will acquire high weights in the final model.Discriminant Analysis is better method of classification. This will produce more accurate and less complex models [8-9].

Conclusion

The algorithm is checked for 50,100,150 and 200 ground truth points and the corresponding overall accuracies and different quality parameters are calculated. The overall accuracy values are obtained are all similar almost in each case. The theme of usage of the different points was to check how far the values are accurate. All the four sets have given similar values in the accuracies and different quality parameters. Generally, with less number of validation points the judgement of a procedure is not correct. In this connection we have verified for different set of validation points. So, it can be concluded that our validation points are correct to use for further procedures.

References

[1]S.Nithyakalyani,B.Gopinath.Analysis of Node    Clustering Algorithms on Data Aggregation in  Wireless Sensor Network.J SciInd Res,vol. 75, 38-
42, 2016.

[2]S J. Lee,T R Lin Grace,P H. His.Industrial Cluster Development and its Contribution to Economic Growth in Taiwan – HsinchuScience and
Industrial Park (HSIP)
.J SciIndRes, Vol.76, 273-
278, 2017.

[3]K. Netti,Y. Radhika. A Model for Accurate Prediction in GeoRSS Data Using Naive Bayes Classifier.J SciInd Res, vol. 76, 473-476,2017

[4] SA. Mousavi, Z. Hanifeloo, P. Sumari, MRM. Arshad.Enhancing the Diagnosis of Corn Pests using Gabor Wavelet Features and SVM Classification.J SciInd Res, 75, 349-354, 2016.

[5] Ch. Swapna,VV. Swetha Kumar,JVR.Murthy.A New Approach to Cluster Datasets without Prior.Knowledge of Number of Clusters, J SciInd Res, vol. 74, 261-264, 2015.

[6] S.Abdikan, F.B. Sanl, M. Ustuner, F. Calo. Land Cover Mapping using Sentinel-1 SAR Data. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, volume XLI-B7,pp.12–19,2016.

[7] N. Alajlan, N. Ammour, Y. Bazi, H. Hichri.A cluster ensemble method for robust unsupervised classification of VHR remote sensing images. in Proc. IEEE IGARSS, pp. 2896–2899, IEEE, 2011.

[8] G. Suresh, R. Gehrke, T. Wiatr, M. Hovenbitzer.Synthetic Aperture Radar Based Classifiers for Land Applications In Germany. The International Archives of the Photogrammetry, Remote Sensing and   Spatial Information Sciences, vol. XLI-B1, pp. 12–19,  2016.

[9] M. Chi, Q. Qian, J. A. Benediktsson. A Survey on SAR Image Classification.International Journal of Advanced Engineering and Global Technology, vol. 3, Issue 12, pp. 1461–1465, 2015.

[10] D. Chen, D. Stow, P.Gong. Examining the effect of spatial resolution and texture window size on classification accuracy: An urban environment case. International Journal of Remote Sensing, Vol. 25, Issue. 11, 2177–2192, 2004.

[11] K. Bahirat, F. Bovolo, L. Bruzzone, S. Chaudhuri. A novel domain adaptation Bayesian classifier for updating land-cover maps with class differences in source and target domains. IEEE Trans. Geosci. Remote Sens., vol. 50, Issue .7, pp. 2810–2826, Jul. 2012.

[12] I.A. Rizvi,B.K. Mohan. Object-based image analysis of high- resolution satellite images using modified cloud basis function neural net- work and probabilistic relaxation labeling process. IEEE Trans. Geosci. Remote Sens., vol. 49, Issue .12, pp. 4815–4820, Dec. 2011.

928 total views, no views today

Download PDF File

About the author: admin