利用密度聚类支持向量机的气象云图云检测

Meteorological imagery cloud detection using density clustering support vector machine

为了提高气象云图云检测的判识精度和计算效率,提出一种基于密度聚类支持向量机(DC-SVM)的云检测方法。分析了MTSAT气象云图的特征提取和选择方案,建立了云和下垫面的分类样本集;在SVM学习中,通过引入样本集的纯度及充足度,选择关键样本,减少了噪声和异常样本的干扰,从而降低了计算复杂度,提高了分类精度。实验表明,该算法的分类正确率较BP神经网络及传统SVM的方法分别提高了2.54%和0.21%,训练时间及测试时间也明显减少;而且,该方法还克服了传统云检测方法需要根据先验知识确定阈值的缺点,检测结果与人工解译结果基本吻合。

In order to improve the recognition accuracy and computation efficiency in cloud detection to meteorological imagery,a cloud detection method based on density clustering support vector machine（DC-SVM） is proposed.The scheme of feature extraction and selection for meteorological imagery is analyzed,and the classification sample set which discriminates cloud form underlying surface is established.In the learning stage of SVM,the critical training samples are selected to reduce the negative impact of noises and outliers by introducing the purity and sufficiency level of sample set,which can reduce the computational complexity and improve the classification accuracy.Experiments demonstrate that compared with BP neural network and traditional SVM,the proposed approach improves the classification accuracy with 2.54% and 0.21% respectively,meanwhile,the training time and testing time is reduced significantly.What is more,the proposed approach overcomes the drawback of traditional cloud detection approaches which depend on prior knowledge to determine the thresholds;the detection result is consistent with the manual interpretation roughly.