论文部分内容阅读
支持向量数据描述SVDD(Support Vector Data Description)在单目标土地覆盖信息提取中得到了初步的应用,但已有研究一般选择纯净像元作为分类样本,不符合SVDD分类器原理,难以保证精度。本文选择北京市通州区为实验区,基于QuickBird重采样图像和地表真值数据构建不同空间特征的训练样本集,设计两种分类样本选择方案:在训练样本集中随机选择分类样本以及手工选择分类样本,进行SVDD分类。实验结果表明,随着惩罚因子C值逐渐增大或核宽度s值逐渐减小,超球会逐渐减小,生产者精度逐渐降低而用户精度逐渐提高。在最优分类参数的前提下,对样本进行SVDD分类。在模拟实验中,随着样本从中心向边缘逐渐扩散,超球也逐渐增大,分类精度逐渐增高,而最外层的样本接近小麦真实分布的边界,超球内混入其他的地物像元;在真实实验中,中心样本集SVDD分类总精度和生产者精度分别为80.48%和45.15%,而边缘样本集SVDD分类总精度和生产者精度分别为92.71%和95.81%,得到了明显的提升(Z=156.12),表明利用边缘像元进行SVDD分类能够提高目标地物的识别精度。
Support Vector Data Description (SVDD) has been initially applied in the single-target land cover information extraction. However, it has been studied that the pure pixels are generally selected as the classification samples and the SVDD classifier does not conform to the principle and it is difficult to ensure the accuracy. This paper chooses Tongzhou District in Beijing as experimental area, constructs training sample sets with different spatial features based on QuickBird resampling images and surface truth values, and designs two kinds of classification sample selection schemes: randomly selecting the classification samples and selecting the classification samples by hand in the training sample set , SVDD classification. Experimental results show that as the penalty value of C increases or the kernel width decreases gradually, the hypersphere will decrease gradually, the accuracy of producer gradually decreases and the user accuracy gradually increases. Under the premise of optimal classification parameters, the samples were classified by SVDD. In simulation experiments, the hypersphere gradually increases as the sample spreads from the center to the edge, and the classification accuracy gradually increases, while the outermost sample is close to the true wheat distribution boundary, and the hypersphere is mixed with other ground object pixels In the real experiment, the total accuracy of the SVDD classification and the producer precision of the central sample set are 80.48% and 45.15% respectively, while the total precision and producer precision of the SVDD classification of the edge sample set are 92.71% and 95.81% respectively, which has been significantly improved (Z = 156.12). It shows that the SVDD classification using edge pixels can improve the recognition accuracy of the object.