词袋模型在高分遥感影像地物分类中的应用研究
2020-09-21王小芹张志梅邵烨王常颖张小峰
王小芹 张志梅 邵烨 王常颖 张小峰



摘 要: 高分辨率遥感影像空间信息丰富,同时也给地物分类带来挑战。故提出一种基于词袋模型的地物分类方法,通过实验讨论词袋模型在这一问题中的适用性。首先在多尺度影像下随机选取场景,通过场景的底层特征聚类建立多尺度视觉词典;然后用视觉单词表达少量标记样本来训练支持向量机;最后用分类器提取典型地物。结果表明,在多尺度词袋模型表达下,研究区分类总体精度达到92.18%,Kappa系数为0.880 9。对比实验结果表明,词袋模型和多尺度词袋模型可以有效表达语义特征,从而在少量标记样本下提高分类精度。
关键词: 高分遥感影像; 词袋模型; 地物分类; 视觉词典; 地物特征提取; 样本表达
中图分类号: TN911.73?34; TP75; TP391 文献标识码: A 文章編号: 1004?373X(2020)17?0056?04
Abstract: High?resolution remote sensing images that contain rich spatial information bring about great challenges to classification of the ground feature. In this paper, a terrain classification method based on the bag of visual words (BOVW) is proposed. The multi?scale visual dictionary is built by clustering of the low?level features in a scene that are randomly selected in the multi?scale image. The visual words are used to express a few marked samples for training the support vector machine. Finally, a classifier is used to extract the typical object features. The results show that, with the expression of the multi?scale BOVW, the overall accuracy of the classification in the study area reaches 92.18%, and the Kappa coefficient is 0.880 9. The comparative experiment results indicate that the BOVW and the multi?scale BOVW can effectively express the semantic features, thus the accuracy of the classification can be improved with a few marked samples.
Keywords: high?resolution remote sensing image; BOVW; ground object classification; visual dictionary; ground object feature extraction; sample expression
0 引 言
2013年以来,我国相继发射了高分系列卫星影像,其中,高分二号的发射意味着我国进入“亚米级”高分时代,高分辨率的遥感影像虽然提供了丰富的地面细节信息,却很容易产生“同物异谱,同谱异物”的现象。传统面向像元的方法已无法解决地物分类问题,基于场景的分类方法成为研究热点。如文献[1]融合场景的像素一致性信息、空间信息和外观信息,采用K?means聚类进行量化并统计直方图,对场景的特征进行表达,实现了高分辨率遥感场景分类;文献[2]提出了一种基于场景的自动识别高铁沿线建筑物隐患目标的方法。其中,基于监督学习的方法可以有效地进行影像地物分类,但是这类方法需要大量标记样本进行训练,而获取大量标记样本需要耗费巨大的人力和物力。……
