基于蚁群算法面向对象的遥感图像分类方法研究

Remote sensing image classification is one of the most fundamental problems in area of remote sensing information analysis and application. The traditional pixel-based classification methods, which utilize only spectral information, can not overcome the main limitation which does not make use of space information in the process of classification.In contrast,object-oriented classification techniques of remote sensing image, integrating information of spectrum, shape and texture of image, utilize spatial information so that to overcome the drawback of the traditional methods while dealing with classification of remote sensing image. However,currently, there is not robust multiscale segmentation algorithm for remote sensing image because of its property of multiscale and complex . Moreover,a single classifier often can not achieve a stably ideal result while dealing with those features beacase object-oriented classification needs to use a variety of high-dimensional features with significant differences in nature and high-dimensional features will lead to curse of dimensionality. Hence,this reseach proposes to apply object-oriented technique fused with multiple classifier based on ant colony optimiaztion algorithm to tackle the problems mentioned above.The main idea of this research is illustred as below. First,the remote sensing images are segemented in multi-scale by using of spectral graph theory and multiple object-oriented features based on spectrum, texture and shape of subimages are extracted from the segmented images with the optimal scale .Secondly,ant colony optimization algorithm combined with genetic algorithm is employed to select the optimal feature subset for members classifier . Lastly,a multiple classifier system model is designed as extension of the new co-evolutionary technique based on ant colony optimization algorithm combined with genetic algorithm, and the construction of multiple classifier system is completed by ant colony optimization algorithm algorithms combined with genetic algorithm , which determine what optimized feature subspace each sub-classifier should use and deal with fusion of multiple classifier as a combinatorial optimization problem,thus, it will be able to effectively improve classification accuracy of object-oriented remote sensing image with high-dimensional features.

遥感图像分类是遥感信息分析与应用中最基本的问题之一。传统基于像素的分类方法难以克服分类过程中未利用空间信息的局限性。面向对象的分类技术利用了空间信息，弥补了传统方法的不足。然而由于遥感图像的多尺度性和复杂性，目前并没有鲁棒的遥感图像多尺度分割算法；其次，由于面向对象的分类方法用到了高维特征，会导致维数灾问题，且单一分类器在处理这些特征时常达不到理想的分类效果。 本项目提出基于蚁群算法和面向对象技术，并融合多分类器系统解决上述问题。首先利用谱图理论对遥感图像进行多尺度分割，并在最优尺度分割图像上提取基于光谱、纹理、形状等多种子图像特征；然后利用蚁群算法融合遗传算法对各成员分类器进行最优特征子集选择；最后，利用蚁群算法融合遗传算法的协同进化新方法对多分类器融合集成学习问题进行优化求解；在已获得各子分类器最优特征子集的基础上，完成多分类器系统的构建，从而有效地提高面向对象遥感图像分类方法的精度。

本项目在分析蚁群算法与遗传算法优缺点的基础上，利用遗传算法改进了蚁群算法性能，提出了一种杂交水稻优化算法，和改进蚁群算法一起为高分辨率遥感图像分析优化应用提供新理论和新方法。针对遥感图像容易出现降质的问题，提出了基于群集智能算法的图像自适应增强算法。针对传统的图像多阈值分割问题，提出了基于群集智能算法优化的多阈值分割方法。针对遥感图像地物的多尺度特征，利用基于谱图理论分割方法对遥感图像进行多尺度分割，基于多尺度MRF模型，提出一种新的高分辨率遥感图像最优分割尺度自动选取算法。该算法将区域特征引入对象，通过对象之间的消息传播，自底向上逐步获取影像的分层结构，最后通过分别计算影像内每个尺度对象的边缘概率极值自动获取最优分割尺度。纹理特征是遥感图像的关键特征，提出并实现了基于改进蚁群算法的图像最优“Tuned”纹理模板训练方法，同时研究和实现了共生矩阵，Gabor小波变换，方向梯度直方图, 局部二阶矩等常用的纹理特征提取方法，进一步研究和实现了遥感图像Zernike矩形状特征提取方法。在进行图像分类的时分类器的性能主要受所采用特征集合以及分类器参数的影响，不同的特征集合和参数往往导致不同的分类结果。针对特征选择问题，提出了一种基于蚁群算法和遗传算法的融合的特征选择方法，可以选择紧凑而有具有良好识别性能的特征子集。对于遥感图像分类器，重点研究了支持向量机和极限学习机的参数优化问题，提出使用改进蚁群算法同步优化支持向量机的特征子集和参数，以获得整体分类性能的支持向量机，实现了原型分类系统；杂交水稻算法优化的极限学习机也具有良好的分类性能。最后利用改进蚁群算法对基于支持向量机和朴素贝叶斯的多分类器分类系统进行优化，实验结果表明优化后多分类器系统能够进一步提高分类器的准确率。本项目的研究，达到了原定的科研目标，有助于提高遥感图像分类的效率和准确率，为遥感图像处理和分析提供理论支持和科学方法。