首页 > , Vol. , Issue () : -
近年来，集成学习成为高光谱遥感影像分类的研究热点，尤其是动态集成算法根据测试样本的特征自适应地选择最佳分类器，其分类性能显著提升。然而现有的动态集成方法仅考虑测试样本与验证样本的光谱信息，忽略了高度规则化的高光谱遥感影像包含的丰富空间信息。为进一步提升高光谱遥感影像动态集成算法分类的准确性和可靠性，提出了联合空间信息的可变K邻域动态集成算法(Variable K-neighborhood and Spatial Information，VKS)和联合自适应邻域空间信息的可变K邻域动态集成算法(Variable K-neighborhood with Shape-Adaptive，VKSA)。两种算法第一阶段综合考虑分类器精度与相似度自适应地改变测试样本的K邻域，第二阶段分别设计固定窗口和自适应窗口的嵌入方式增加地物的局部空间近邻关系，充分利用高光谱遥感影像地物复杂的空间形态结构信息。实验部分采用三组通用的高光谱遥感影像数据对所提出算法的性能进行综合评价。结果表明相比于传统的动态集成算法，本文提出的联合空间信息的动态集成模型能显著提升分类精度，其中基于自适应窗口方式的VKSA算法明显优于基于固定窗口的VKS算法。
Recently, ensemble learning has attracted much attention for hyperspectral image analysis. It is the model of in-tegrating multiple base classifiers to jointly make decisions, which are deemed to be better than a base classifier. Ensemble learning includes static classifier ensemble and dynamic classifier ensemble. In the static ensemble method, the same classifier combination scheme is selected for the classification of testing sample. However, this method ignores the difference of classifier performance for each testing sample. Considering the features of test-ing sample, the best classifier is selected adaptively in dynamic ensemble methods. So it generally can achieve better performance than static ensemble methods for hyperspectral image classification. However, a lot of dy-namic ensemble methods only consider the spectral information of validation and training sample, ignoring the fact that hyperspectral image also contains rich spatial information. In order to further improve the accuracy and reliability of hyperspectral image classification, a variable K-neighborhood and spatial information algorithm (VKS) is proposed in this paper. Firstly, the VKS algorithm considers the accuracy and similarity of classifier comprehensively to adaptively adjust the K neighborhood of testing sample, which makes the setting of Region of Competence (RoC) more reliably and flexibly. Thus, the testing sample with good spectral discrimination performance are classified preferentially. For the testing sam-ples with poor spectral discrimination performance, the label information of spatial neighborhood samples is used for prediction. A fixed window is designed to provide local spatial information of hyperspectral image. However, fixed windows can not reveal the complex and changeable morphological characteristics of ground objects. In order to capture the complex and changeable spatial structure in hyperspectral image, an adaptive window is proposed which can better reflect the complex spatial information, a variable K-neighborhood with shape-adaptive algorithm (VKSA) is further designed. Purdue Campus, Indian Pines and Salinas hyperspectral remote sensing data are used to design experiments and testify the performance of the proposed VKS and VKSA. Four state-of-the-art methods, namely, majority voting(MV), overall local accuracy (OLA), modified local accuracy (MLA), and multiple classifier behavior (MCB), are used to quantify the classification accuracy. Experimental results demonstrate the VKS and VKSA outperforms static ensemble methods and three classic dynamic ensemble methods in overall classification ac-curacy. Moreover, the VKSA algorithm with the adaptive window can provide better performance than the VKS algorithm with the fixed window.