论文部分内容阅读
目的高光谱遥感影像由于其巨大的波段数直接导致信息的高冗余和数据处理的复杂,这不仅带来庞大的计算量,而且会损害分类精度。因此,在对高光谱影像进行处理、分析之前进行降维变得非常必要。分类作为一种重要的获取信息的手段,现有的基于像素点和图斑对象特征辨识地物种类的方法在强噪声干扰训练样本条件下精度偏低,在对象的基础上,将光谱和空间特征相似的对象合并成比其还要大的集合,再按照各个集合的光谱和空间特征进行分类,则不容易受到噪声等因素的干扰。方法提出混合编码差分进化粒子群算法的双种群搜索策略进行降维,基于支持向量机的多示例学习算法作为分类方法,构建封装型降维与分类模型。结果采用AVIRIS影像进行实验,本文算法相比其他相近的分类方法能获得更高的分类精度,达到96.03%,比其他相近方法中最优的像元级的混合编码的分类方法精度高出0.62%。结论在针对强干扰的训练样本条件下,本文算法在降维过程中充分发挥混合编码差分进化算法的优势,分类中训练样本中的噪声可以看做多示例学习中训练包“歧义性”的特定表现形式,有效提高了分类的精度。
Purpose Hyperspectral remote sensing images, due to their huge number of bands, lead directly to the high redundancy of information and the complexity of data processing, which not only brings huge computational load but also impairs classification accuracy. Therefore, it is necessary to reduce the dimension before processing and analyzing hyperspectral images. Classification as an important means of obtaining information, the existing methods of identifying feature types based on the characteristics of pixel points and patch features are less accurate under strong noise interference training samples. Based on the objects, the spectral and spatial Objects with similar characteristics are merged into a larger set, and then classified according to the spectral and spatial characteristics of each set, the noise is less susceptible to interference. Methods A two-population search strategy based on hybrid-coded differential evolution particle swarm optimization algorithm is proposed to reduce the dimension. Support vector machine based multi-sample learning algorithm is used as classification method to construct encapsulated dimensionality reduction and classification models. Results AVIRIS images were used for experiments. Compared with other similar classification methods, the proposed algorithm can achieve higher classification accuracy of 96.03%, which is 0.62% higher than the best pixel-level hybrid coding classification method in other similar methods. . Conclusions Under the condition of training samples with strong interference, the proposed algorithm takes full advantage of the hybrid encoding differential evolution algorithm in the dimensionality reduction. The noise in the training samples can be regarded as the training package "ambiguity The specific form of expression, effectively improve the classification accuracy.