1 semantic geo-image classification using image ... semantic geo-image classification using image...

Download 1 Semantic Geo-Image Classification Using Image ...  Semantic Geo-Image Classification Using Image Processing Techniques Annai College Of Engineering  Technology Kumbakonam Department of Information Technology Guided by:Mrs.S.vidivelli ME.,Phd.,(AP/IT). Prabha.M

Post on 15-Apr-2018

214 views

Category:

Documents

2 download

Embed Size (px)

TRANSCRIPT

  • 1

    Semantic Geo-Image Classification UsingImage Processing Techniques

    Annai College Of Engineering & Technology KumbakonamDepartment of Information TechnologyGuided by:Mrs.S.vidivelli ME.,Phd.,(AP/IT).

    Prabha.M(prabhampps1994@gmail.com ) , Sabitha.M(sabithajayakaran@gmail.com) ,Vinothini.K(lathavinothiniit@gmail.com) ,

    Abstract:

    Satellite image classification is one of the most significant applications

    in remote sensing. Remote sensing data obtained from different optical

    sensors have been commonly used to characterize and quantity land

    information. However, conventional optical remote sensing is limited

    by weather conditions. Synthetic aperture radar (SAR), with the all-

    weather and all-time advantages, is important in the domain of earth

    observation. Polarimetric SAR (PoISAR) images can provide more

    target information and facilitate improvement of the land cover

    classification accuracy. Therefore, land cover classification for

    PolSAR images is important in remote sensing, especially for those

    areas that change drastically with season. This project implement

    algorithm for the land classification for the PoISAR images. We can

    propose multilevel semantic features approach to extract the high level

    features such as Entropy/Anisotropy/Alpha values. And implement

    physical scattering properties and implement Latent Dirichlet

    allocation scheme to discover high level semantics to provide

    histogram for each pixels. Finally implement KNN classification to

    classify the PoISAR images with various class labels such as water,

    land and other properties. Experimental results validate the feasibility

    of the proposed method for land cover classification of the various

    places, Le., the overall accuracy reaches up to 90.91 %, while that for

    the method based on the Wishart distance is 85.01 %, which exhibits

    the superiority of the proposed method over state of art classification in

    the various geo spatial data.

    I. INTRODUCTION

    HE multitude of modern remote sensing sensors allowsus to analyze tremendous amounts of high-resolution

    earth observation (EO) images. Therefore, developing newcontent-based image retrieval (CBIR) systems being able toextract user-desired information from existing image data-bases is highly demanded. In the state-of-the-art literature,various CBIR systems have been proposed for EO imagemining, such as Intelligent Interactive Image KnowledgeRetrieval System (I3KR) [1], Knowledge-Driven InformationMining System [2], and its accelerated variant [3]. For theexisting systems, semantic image interpretations are usuallyprovided through either manual image annotation or useracceptance (in active learning scenarios), which require muchhuman effort and time, and bias the systems toward user

    Fig. 1. Stepwise low- to high-level semantics generation.

    perspectives [4]. Moreover, due to the differences betweenhuman image understanding and how computers interpretand process them (the so-called semantic gap [4], [5]),many image mining results provided by computers are stillunsatisfactory.

    In this letter, we propose a multilevel semantics discoveryapproach for bringing computer interpretation of a particularremote sensing image type, namely, polarimetric syntheticaperture radar (PolSAR) images, closer to human semantics.This helps computers to discover existing semantic relation-ships within images and employ them for analyzing userqueries (which are based on semantics) and provide themwith semantically meaningful relevant results. Fig. 1 showsan overview of the proposed approach.

    PolSAR images supply information with respect to thephysical scattering properties of the recorded ground targets,retrieved by applying coherent or incoherent target decom-position theorems to the first- and second-order polarimetricrepresentations [6]. A widely employed decomposition methodis Entropy/Anisotropy/Alpha (H/A/), resulting in three para-meters describing the physics behind the scattering processes.These parameters lead to a superior pixel-based unsupervisedclassification scheme, the H/A/ classification method [7].A modification to this method, namely, the H/A/-Wishartclassification method [8], shows that the complex Wishartdistribution parameters improve the classification substantially.

    In our proposed approach, we employ the H/A/-Wishartmethod for discovering the low-level semantics of PolSARimages as a set of classes, representing targets by their physicalscattering properties (e.g., low-entropy surface scattering (SS)with high anisotropy). The images are then tiled into patches,and each image patch is modeled as a bag-of-words (BoW)by generating a histogram of its assigned class labels, wherethe labels are the words in the BoWs (see Fig. 2). Afterthat, a generative statistical model, namely, latent Dirichletallocation (LDA) [9], is applied to the BoW histograms inorder to discover the latent semantics behind the image patches

    vts-6Text Box

    vts-6Text BoxISSN: 2348 - 8387 www.internationaljournalssrg.org Page 76

    vts-6Text BoxSSRG International Journal of Computer Science and Engineering- (ICET'17) - Special Issue - March 2017

  • 1

    Fig. 2. H/ classification scheme (adapted from [7]). From the 18 classes,only 116 are feasible and are assigned as labels L1L16 in Table I.as a set of topics. Our validation of the topics based on groundtruth (Google Earth1) images demonstrates that they provide Fig.3.grayscale conversion median filter

    From (1) and (2), the following parameters can be computedhigh-level semantics, which are close to human semantics usedfor identifying land-cover types (e.g., woody vegetation).

    As the topics are the basic land-cover types existing in

    3

    H = pi log3 pi , A =i=1

    p2 p3p2 + p3

    3

    , = pii (3)i=1

    the image patches, further land-cover types can be defined as where pi = i / 3 j denotes the probability of thecombinations of the basic land-cover types (e.g., a shorelineis a combination of a water body, grassland, and woody vege-tation). Thus, the image patches can be modeled by vectors oftopic mixtures, the co-called bag-of-topics (BoT) model [10].The topic vectors form a multidimensional Euclidean spacein which each image patch is represented as a point. Sincethe dimensions of this space are semantically meaningful,it can be easily explored and assessed through immersivevisualization techniques to discover existing semantic rela-tionships and identify new semantic categories (e.g., mixedland-cover types). While a topic representation of images canadapt computer image interpretation to human semantics, thesemantic relationships can be used for designing rule-basedland-cover categorization methods.

    Section II reviews H/A/-Wishart classification. Section IIIdescribes low-level semantics discovery. Section IV brieflyintroduces LDA. Sections V and VI explain how to discoverhigh-level semantics and explore it for finding new semantics.Section VII concludes this letter.

    II. ENTROPY/ANISOTROPY/ALPHA-WISHARTCLASSIFICATION

    Our test image is an F-Synthetic Aperture Radar System air-borne complex-valued data set2 comprising four polarizationplanes (VV, HH, HV, and VH) [11]. The data were multilookedwith a factor of 5 in azimuth direction. In order to reducethe inherent speckle noise, PolSAR data are usually deliveredin a multilooked pixel-wise coherency matrix format T . Thismatrix is a 3 3 Hermitian, positive, and semidefinite matrix,which can be written as

    T = U U1 (1)

    where is a diagonal matrix composed of the eigenvaluesof T (1 , 2 , and 3 in descending order). The columns ofU contain the corresponding eigenvectors to the eigenvalues(u1, u2 , and u3), where each ui can be further decomposedinto [6]

    u i = cos i sin i cos i e j i sin i cos i e jiT

    . (2)

    1https://www.google.com/earth/2courtesy of Rolf Scheiber from DLRs Microwaves and Radar Institute

    eigenvalues i . These parameters refer to the physics behindthe scattering processes. The entropy H discriminates pure anddistributed scatterers; the anisotropy A characterizes differenttypes of scattering [6], and the mean angle shows thedominant scattering mechanism.

    Combinations of these parameters can lead to very goodclassification schemes. For example, the scheme in Fig. 2divides the H / plane into nine zones (classes), from whichonly eight are feasible. Including anisotropy will double thenumber of classes: nine for A 0.5 (denoted by odd labels)and nine for A > 0.5 (denoted by even labels). The H/A/classification method has its own drawbacks, such as the fixedboundaries of the classes, which do not dynamically adapt tothe input data. Dealing with this issue, Lee et al. [8] proposedto use the parameters of the complex Wishart distribution ofthe coherency matrix. This approach employs the H/A/ clas-sification for initializing the classes, and then uses the Wishartparameters in an iterative procedure to refine the classification.Its main advantages are to consider the magnitude of thecoherency matrix, which is important in detecting SS, andallowing a dynamic adaptation of the class boundaries to theinput data.

    III. LOW-LEVEL SEMANTICS DISCOVERY

    We apply the H/A/ Wishart classification, based on5 5 pixel windows, to our PolSAR test image, shown inFig. 3 (left), in order to retrieve a classification map with16 classes. The class labels are then considered as low-levelsemantics, which refer to the scattering properties of therecorded targets. Fig. 3 (right) shows the classification results,with Table I defining the color coding.

    The labels in Table I can be categorized into three mainscattering mechanisms: the first six labels (L1L

Recommended

View more >