优化选择的融合纹理—光谱—空间相关性多特征一体化的高光谱图像分类方法研究

Hyperspectral imagers can proivide detailed spectral information of various ground cover types due to its wide coverage of wavelength and high sampling rate.It is very important to make use of hyperpspectral data for target recognition and landscapes classification. However,it is difficult to deal with higher spectral variance within each class corresponding to land-cover units only using spectral features.The improtance of integration of texture, spatial contexture and spectral patterns simultaneously has been identified as a desired goal by many scientists devoted to hyperspectral data analysis.However, there exits three issures in spatial-spectral classification:(1)Individual texture extraction method can not handle reqirement of complex texture.(2)It is very lack of classificaiton discriminant model which can integrate texture, spectal and contextrual features simultaneously.(3)Class boundary and details of structure information has been ignored easily in classification process.Address these issues, a method that integrates texture, spectral and spatial contextual features for improved hyperspectral classification is presents.At the beginning, proper fusion of texture features derived from different texture method is expected to produce an improved feature set based on Intelligent ant colony optimal (ACO) algorithm.Next, the extended Markov random field random (extended-MRF) has been proposed to integration of texture, spectrum and contexure multiple features. We expect to achieve accuate classification map of hyperspectral images. The research results will increase the overall accuracy, at the same time, keep class boundaries and the details of structure information in the process.

高光谱遥感能够提供丰富的地物波谱信息，利用高光谱图像进行地物目标识别和分类具有重要意义。仅利用光谱特征分类难以解决光谱异质性造成的"同物异谱"等问题，将纹理特征或空间相关性特征引入高光谱图像分类中，已成为近年来研究的热点，但仍存在如下问题：（1）单一纹理特征提取方法已难以满足不同地物类型复杂纹理提取的需求；（2）缺乏能够同时将纹理特征、光谱及空间相关性特征等多特征整合的分类模型；（3）忽视了地物边界等细节结构信息的保持。针对以上问题，本项目将以纹理-光谱-空间相关性多特征一体化为核心研究高光谱图像精细分类。首先，基于不同纹理提取算法结果，展开基于智能化蚁群算法的多纹理特征融合算法研究；其次，提出基于扩展马尔科夫随机场的纹理-光谱-空间相关性多特征一体化分类判别模型。通过本项研究，实现高光谱图像精细分类，在提高整体分类精度的同时保持地物边界细节信息

本项目完成了既定科学目标，研究了高光谱图像空间特征的提取方法，特别是空间上下文特征和纹理特征，将优化的空间纹理，空间相关性以及光谱特征进行融合，通过整合智能化算法及改进型马尔科夫随机场模型，较好的提高了高光谱图像的分类精度。而且，将该特征提取及融合方法扩展至高光谱图像与LiDAR数据的融合分类中，得到了更好的效果。在此基础上将本项目的算法进行实际应用，如基于卫星影像的北京地区土地覆盖分类方法以及长时间序列的地物变化检测。本项目支持发表SCI论文4篇，另有7篇文章发表在CSCD核心期刊上。