Gravity Model Based Approaches For Effective Classification Of High Resolution Remote Sensing Imagery

The development of high-spatial-resolution（HSR）remote sensing has facilitated the extraction and analysis of the ground information with fine detail.Classification of HSR imagery,as a critical step for information extraction,has attracted extensive attention.Due to vast volume of data and abundant information generated under complex and varying spatial-temporal and spectral conditions,there is still lack of in-depth and systematic solution for feature extraction and image classification.In fact,both these challenging tasks including feature extraction and data classification can be intrinsically converted to certain optimisation problems.To tackle these problems,theoretically meta-heuristic search algorithm has provided possible solutions owing to its superior self-organization,self-adaptation and global optimization abilities.Inspired by the gravity model and the gravitational search algorithm（GSA）,a set of novel approaches are proposed for effective classification of HSR imagery.Through the definition of the corresponding objective functions,selection of features and optimization of classifiers are achieved.The major contributions of this thesis are highlighted as follows.（1）A gravitational search algorithm based on dynamic neighborhood learning-based GSA（DNLGSA）is proposed based on the deeply analysis of the optimization mechanism of GSA.It was found that the K_bestest model in conventional GSA is a global model which suffers from weak local search capability.To overcome this problem,a locally full-connected neighbourhood model is proposed to replace the K_best model for improved local search capability.Moreover,to further improve the self-adaption capability of GSA,we propose a dynamic neighbourhood learning（DNL）scheme to dynamically construct the neighbourhoods based on convergence states of the algorithm.In DNLGSA,the convergence state is delineated by two convergence criteria named limit value and population diversity.Obviously,this method overcomes the insufficient information exchange between local neighbourhoods whilst keeps the diverse search direction property of GSA.The experimental results reveal that DNLGSA exhibits competitive performance when compared with a variety of state-of-the-art meta-heuristic algorithms.Moreover,the time-consuming of DNLGSA is significantly less than other GSA variants.（2）A data field-based multi-objective gravitational search algorithm（DFMOGSA）is proposed.In DFMOGSA,external archives are taken as the data field in the objective space,which are applied to store the obtained non-dominated solutions.Accordingly,each non-dominated solution is assigned a density value based on its potential energy.The solution with the smallest density value is chosen as the first kind of guide-particle.This guide-particle directly leads population particles to convergence towards the low-density region and thus improves the solution distribution.Moreover,several superior population particles are selected as the second kind of guide-particles.These guide-particles direct each of the population particles to fully explore the feasible search space via their resultant gravitational force,which ensures the convergence performance.Simulation results on several benchmark test problems show the effectiveness and superiority of the proposed DFMOGSA for multi-objective optimization problems.（3）On the basis of the DNLGSA,an information extrication and feature selection method is developed.First,the spectral and texture-based spatial features are extracted from the HSR imagery to form the preliminary feature set.The DNLGSA is then mapped to the binary space with each dimension representing a feature.Classification accuracy from the support vector machine（SVM）and the number of selected spectral bands are utilized to measure the discriminative capability of the feature subset.Finally,the feature subset with the smallest number of features which covers the most useful and valuable information is obtained.The experimental results have shown that the proposed method can indeed considerably reduce data storage costs and efficiently identify the feature subset with stable and high classification accuracy.（4）A multi-lvevel thresholding approach for multi-feature HSR imagery is constructed based on the DNLGSA.The main object of image segmentaion is to divide an image into different regions of different size.The characteristicswithin a regionis is similar while quite different in different regions.This method can take full advantage of the spacial information in HSR imagery and make up the deficiency of simple classification method.Thresholding method is simple and effective segementeion method with wide application.However,when dealing with complicated images,it easily suffers from premature convergence problem.Moreover,the tradition Second,traditional methods thresholding is developed for gray image segmentation which is not applible for the multi-feature HSR imagey.To overcome these propoblems,the proposed method establishes the space mapping of multi-features and thresholding by combining the DNLGSA and Otsu criterion.Then the optimal thresholds are generated by the optimization of Otsu criterion.The experimental results reveals that the proposed method can overcome the timie-sonsuming and premature convergence problems in traditional multi-level thresholding algorithms.（5）A multi-objective image classification method for HSR imagery is designed based on the DFMOGSA.Simultaneous optimization of different cluster validity measures can capture different data characteristics of HSR imagery and thereby achieve high accuracy classification results.Thereby two conflicting cluster validity indices,the Xie-Beni（XB）index and Jm measure,are integrated to realize better classification of the HSR imagery on the basis of the DFMOGSA.The obtained classification results are compared with that are produced by two single cluster validity indices based approaches and two state-of-the-art multi-objective optimization algorithms.Comparison results show that the proposed method can achieve more accurate image classification.