基于字典学习的小样本高光谱遥感图像稀疏表示分类精度研究与应用

项目名称： 基于字典学习的小样本高光谱遥感图像稀疏表示分类精度研究与应用

项目编号： No.61461002

项目类型： 地区科学基金项目

立项/批准年度： 2015

项目学科： 无线电电子学、电信技术

项目作者： 张春梅

作者单位： 北方民族大学

项目金额： 40万元

中文摘要： 针对荒漠、戈壁等无人区标记样本获取困难且成本高的难题，采用基于学习字典的稀疏表示方法，对小样本情况下高光谱遥感影像分类精度问题进行研究。 稀疏表示利用字典的冗余特性捕捉信号本质特征，在统计模式识别领域具有潜在优势。基于学习字典的稀疏表示能够利用少数样本建立训练集，通过字典学习不断更新字典，能较好的适应少标记样本问题。 主要研究内容包括:(1)研究和改进字典自身学习更新的快速算法，提高字典学习的效率和收敛性。(2)研究样本个数与分类精度的关系曲线及其收敛性，研究影响该曲线关系的主要因素和性能指标。(3)构建荒漠植被的光谱、植被和纹理等特征的组合特征向量模型，研究宁夏戈壁荒漠地区植被覆盖标记样本的获取质量和分类精度。 通过该项基础应用研究，促进稀疏表示分类技术更好地走向实际工程应用，为宁夏荒漠、戈壁等无人区植被覆盖标记样本的获取数量和植被分类质量提供理论依据和技术支持。

中文关键词： 遥感图像；图像分类；高光谱分类；小样本分类

英文摘要： The hyperspectral data classification of depopulated zones like desert and gobi is a challenge with small labelling samples because of acquiring samples difficultly and expensively in most situations. So, this study will try to improve the accuracy of hyperspectral classification with small samples by sparse representation based on dictionary learning. Sparse representations are able to extract the nature features of signals with redundant dictionaries and have a great potential in statistic pattern recognition. Sparse representation with dictionary learning can build an initial dictionary with a few training samples to solve the problems of insufficient labelling samples by constantly updating dictionaries. The research contents include：(1) To study the algorithm of dictionary update to improve its efficiency and convergence. (2) To study the functional relationship between classification accuracy and the amount of training samples, as well as the convergence and affecting factors of this funtional curve. (3) To constructing the feature models with multi-spectral features, as spectrum, vegetation and textures. To study the quality of data acquisition for labelling samples and improve classification accuracy in the depopulated zones. With this fundamental application research, classification application of sparse representation can be made progress toward engineering area and provide the theory and technical supports for hyperspectral data classifacation in the depopulated zones in Ningxia province.

英文关键词： remote sensing image;image classifacation;hyperspectral classification;small samples classification