ashoka socpros2012 paper 62

Upload: ashoka-vanjare

Post on 03-Jun-2018

223 views

Category:

Documents


0 download

TRANSCRIPT

  • 8/12/2019 Ashoka SocProS2012 Paper 62

    1/21

    MULTI-TEMPORAL SATELLITE IMAGE ANALYSIS

    USING GENE EXPRESSION PROGRAMMING

    J. Senthilnath , S.N. Omkar , V. Mani , Ashoka Vanjare,P.G. Diwakar

    Ashoka Vanjare SocPros 2012 1

  • 8/12/2019 Ashoka SocProS2012 Paper 62

    2/21

    Outline

    Motivation Objective Problem formulation Methodology

    Study area Results & discussion Conclusions & future work References Acknowledgement Recognitions

    Ashoka Vanjare SocPros 2012 2

  • 8/12/2019 Ashoka SocProS2012 Paper 62

    3/21

    Earth observation sensor namely optical Moderate ResolutionImaging Spectroradiometer (MODIS) image is useful for floodassessment application

    The spectral (pixel) classification is carried out using classificationtechniques.

    Incorporating spatial features with spectral information results ingood accuracy.

    To overcome this problem nature inspired algorithm is used.

    Motivation

    Ashoka Vanjare SocPros 2012 3

  • 8/12/2019 Ashoka SocProS2012 Paper 62

    4/21

    Objective

    Multi-temporal MODIS satellite image analysis using Artificial NeuralNetwork (ANN) and Gene Expression Programming (GEP) classifiers to assessflooded regions.

    Use of image low resolution image about 250 meters square

    Spectral features of an image is used to differentiate water and non-waterimage pixels

    An automatic extraction of river networks (using imagery before floodingand during flooding)

    Evaluating extracted floods results with ground truth data

    Ashoka Vanjare SocPros 2012 4

  • 8/12/2019 Ashoka SocProS2012 Paper 62

    5/21

    Problem formulation

    Identification of similar water image pixels which is obtained bypartitioning the image and grouping similar pixels

    The given problem is solved by grouping of pixels i.e. by imagesegmentation through artificial neural network and Gene expression

    programming methods.

    Ashoka Vanjare SocPros 2012 5

  • 8/12/2019 Ashoka SocProS2012 Paper 62

    6/21

    Methodology

    Here we first classify and then segment

    Image Classification Two supervised classification methods Artificial Neural Network and GEPare used to classify the similar spectral water features of an image

    Training data set are created in order to identify similar pixels.

    During classification, some of water image pixels are misclassification as

    non-water image pixels so it is resolved by using region-growing andknowledge-based techniques

    Ashoka Vanjare SocPros 2012 6

  • 8/12/2019 Ashoka SocProS2012 Paper 62

    7/21

    Methodology For the given training data set, ANN generates weights where as GEP provides

    mathematical expression

    Using these weights and mathematical expression all pixels of the image are

    extracted and evaluated

    Ashoka Vanjare SocPros 2012 7

  • 8/12/2019 Ashoka SocProS2012 Paper 62

    8/21

    MethodologyImage Segmentation Image segmentation is to eliminate those pixels that are wrongly classified aswater image pixels

    By using shape index and density index are used for identifying non linearfeatures like river.

    A

    P SI

    4 )()(1 Y VAR X VAR

    A DI

    Where;P- represents the perimeter of the region

    A -represents the area of the region)( X VAR -represents variance of x -coordinates

    )(Y VAR -represents variance of y -coordinates

    Ashoka Vanjare SocPros 2012 8

  • 8/12/2019 Ashoka SocProS2012 Paper 62

    9/21

    Methodology The river image pixels have a high Shape Index and a low Density Index,whereas non-river pixels have a low Shape Index and a high Density Index

    So a threshold values are set for both indices for segmenting water image

    pixels from non-water pixels

    Ashoka Vanjare SocPros 2012 9

  • 8/12/2019 Ashoka SocProS2012 Paper 62

    10/21

    Study Area

    Ashoka Vanjare SocPros 2012 10

    Figure 1: Study area covers the districts - Kurnool,Mahaboobnagar, Bellary, Gulbarga and Raichur

    The study are is Krishna river in south India and images cover anarea of 75.55 km 2

  • 8/12/2019 Ashoka SocProS2012 Paper 62

    11/21

    Results and Discussions

    Figure 2. (a) MODIS image of Krishna River before flood. (b) Segmented Image usingANN. (c) Segmented Image using GEP

    The two classifier are used to extract the course of the river from theMarch 2009 image (i.e. before the flooding)

    Ashoka Vanjare SocPros 2012 11

  • 8/12/2019 Ashoka SocProS2012 Paper 62

    12/21Ashoka Vanjare SocPros 2012 12

    Fig. 3. (a) MODIS during flood image with ground truth informationflooded (white dots with black dots) and non-flooded cities (white dots). (b) Segmented Image usingANN (White pts. are identified as flooded cities). (c) Segmented Image using GEP (White pts. are

    identified as flooded cities).

    Results and Discussions

    Two MODIS images (i.e. before flooding and during flooding) of the same regionare used to classify, segment and validate.

    The classifiers were trained using 20 randomly picked samples of two classes(water and non-water) and the whole image was tested using the trainedclassifiers.

  • 8/12/2019 Ashoka SocProS2012 Paper 62

    13/21

    Results and Discussions

    Where; B2 represent the Band 2 of the MODISimage.

    5.4)2*2( B M

    )2*3.11(3.3 B N

    If the value of M and N for a pixel is > 0.5 then this pixel is classified aswater; else non-water pixel

    Further misclassified are segmented using region-based segmentationtechnique.

    GEP expression tree generated for the training samples using before andduring flooded image

    Ashoka Vanjare SocPros 2012 13

  • 8/12/2019 Ashoka SocProS2012 Paper 62

    14/21

    Results and Discussions For image segmenting, we have used two indices - Shape Index (SI) andDensity Index (DI)

    These indices are used to differentiate the water image pixels from non-water pixels

    For March 2009 image, we have used a SI threshold value of 2.7 and a DIthreshold value of 0.8

    For September 2009 image, we have used SI threshold value of 2.7 anda DI threshold value of 1.2

    After segmenting by ANN and GEP, we have verified it with ground truthdata

    The performance of GEP is better than ANN in water extraction of both

    before and during flood imagesAshoka Vanjare SocPros 2012 14

  • 8/12/2019 Ashoka SocProS2012 Paper 62

    15/21

    Image RMSE of ANN RMSE of GEP

    Before flood(March 2009)

    0.1260 0.1205

    Features ANN GEP

    True positive rate 0.83 1

    True negative rate 0.94 0.94

    False positive rate 0.06 0.06

    Accuracy 0.89 0.96

    Table 1. RMSE for before flood image obtained by the two classification techniques

    Table 2. Evaluating features based on ROC for during flood image

    Results and Discussions

    Root Mean Square Error (RMSE) parameters is used for segmentation evaluation

    Receiver Operating Characteristic (ROC) parameter is used for extraction evaluation

    Based on the RMSE values for before flood image, GEP classification results inless error in extraction than ANN classification

    GEP produces less error than ANN

    Ashoka Vanjare SocPros 2012 15

  • 8/12/2019 Ashoka SocProS2012 Paper 62

    16/21

    Results and Discussions

    Figure 2. (a) MODIS during flood image with ground truth informationflooded and unflooded cities. (b) Segmented Image using ANN (White pts. areflooded cities). (c) Segmented Image using GEP (White pts. are flooded cities).

    Notrecognizedplace by ANN

    Ashoka Vanjare SocPros 2012 16

  • 8/12/2019 Ashoka SocProS2012 Paper 62

    17/21

    The tasks of river mapping and flood extraction are accomplishedsuccessfully by the procedure of pixel-based spectral information forclassification, and shape information for segmentation.

    In the classification stage of extracting water and non-water groups, thegene expression programming classifier proved better than the artificialneural network classifier.

    The results of classification using spectral information are improvedthrough region-growing image segmentation (based on spatial feature)using similarity criteria emphasizing shape information.

    In future, research work is carried out using high resolution image andradar images

    Conclusions and future works

    Ashoka Vanjare SocPros 2012 17

  • 8/12/2019 Ashoka SocProS2012 Paper 62

    18/21

    [1] Brakenridge, R., Anderson, E.: MODIS-based flood detection, mapping and measurement: The potential foroperational hydrological applications. Transboundary Floods: Reducing Risks through Flood Management Springer-Verlag, 1 12, (2006). doi: 10.1007/1-4020-4902-1[2] Khan, S. I., Hong, Y., Wang, J., Yilmaz, K. K., Gourley, J. J., Adler, R. F., Brakenridge, G. R., Policelli, F., Habib, S., Irwin,D. : Satellite Remote Sensing and Hydrologic Modeling for Flood Inundation Mapping in Lake Victoria Basin:Implications for Hydrologic Prediction in Ungauged Basins. IEEE Transactions on Geoscience and Remote Sensing, 49,85 95, (2011), doi: 10.1109/TGRS.2010.2057513[3] Mingjun, S., Daniel, C.: Road Extraction Using SVM and Image Segmentation, Photogrammetric Engineering andRemote Sensing, 70(12), 1365 1371, (2004)

    [4] Wang, Z., Zhang, D.: Progressive switching median filter for the removal of impulse noise from highly corruptedimages, IEEE Transactions on Circuits and Systems II, 46, 78 80, (1999) doi: 10.1109/82.749102[5] Haykin, S.: Neural Networks - A Comprehensive Foundation, 2E, Pearson Prentice hall publication, (1994)[6] Ferreira, C.: Gene Expression Programming: a New Adaptive Algorithm for Solving Problems. Complex Systems,13(2), 87-129, (2001)[7] C. S. Arvind, Ashoka Vanjare, S. N. Omkar, J. Senthilnath, V. Mani and P. G. Diwakar.: Multi-temporal Satellite ImageAnalysis Using Unsupervised Techniques", Advances in computing and information technology. Advances in IntelligentSystems and Computing, 177, 757-765, (2012) doi: 10.1007/978-3-642-31552-7_77[8] Fawcett, T.: an introduction to ROC analysis. Pattern Recognition Letters, 27(8), 861 874, (2006), doi:10.1016/j.patrec.2005.10.010.[9] J. Senthilnath, Shivesh Bajpai, S.N. Omkar, P.G. Diwakar, V. Mani: An approach to multi-temporal MODIS imageanalysis using image classification and segmentation, Advances in Space Research, 50(9), 1274 1287, (2012) doi:http://dx.doi.org/10.1016/j.asr.2012.07.003[10] Omkar, S.N., Mani, V., Diwakar, P.G, Senthilnath, J., Ashoka Vanjare: Multi-temporal time series analysis of satelliteimage, Technical Report no: AE/ISTC/SNO/10/238/02, IISc, Department of Aerospace Engineering, Bangalore.

    Ashoka Vanjare SocPros 2012 18

    References

  • 8/12/2019 Ashoka SocProS2012 Paper 62

    19/21

    This work is supported by the Space Technology Cell, Indian Institute ofScience, Bangalore and Indian Space Research Organization (ISRO).

    We also acknowledge the MODIS mission scientists and associated NASApersonnel for the production of the remote sensing data which is used inthis paper.

    Ashoka Vanjare SocPros 2012 19

    Acknowledgement

  • 8/12/2019 Ashoka SocProS2012 Paper 62

    20/21

    Ashoka Vanjare SocPros 2012 20

    Recognitions-Related works

  • 8/12/2019 Ashoka SocProS2012 Paper 62

    21/21

    Thank you

    Ashoka Vanjare SocPros 2012 21