混合编码差分进化粒子群算法及多示例学习的高光谱影像降维与分类

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。

Dimension reduction and classification for hyperspectral image based on particle swarm optimization and differential evolution algorithm with hybrid encoding and multiple instance learning

Objective The high dimensions of hyperspectral remote-sensing images cause information redundancy and data processing complexity, thus leading to high computing workloads and low application accuracy. Therefore, before analyzing hyperspectral images, the high dimensions of hyperspectral data should be reduced. Classification is an important means of acquiring information, but existing image classification methods for identifying different ground objects at the pixel level and object level have low accuracies under the condition of strong noise interference to training samples. This interference decreases when similar objects with spectral and spatial characteristics are merged into large collections and classified according to the spectral and spatial characteristics of each collection. Method This paper proposes a double-population hybrid search strategy for dimension reduction based on differential evolution and particle swarm optimization with hybrid encoding. In this strategy, a support vector machine is adopted as a classifier with multiple-instance learning to improve the classification accuracy, reduce the dimension, and construct the classification model of encapsulation type. Result Experiments were conducted with AVIRIS images. Results show that the proposed method can obtain a classification accuracy of 96.03% for small training samples, 0.62% higher than the best classification accuracy among similar hybrid encoding methods of classification. Conclusion Under the strong noise interference of training samples, the proposed strategy utilizes the double-population hybrid search strategy, which is based on differential evolution and particle swarm optimization with hybrid encoding. The noise interference can be viewed as the specific form of "ambiguity" in the training package of multiple-instance learning. The proposed method can obtain high classification accuracy for small training samples and significantly alleviate strong noise interference.