基于马氏距离和SLIC算法的云检测模型

Cloud Detection Model Based on Mahalanobis Distance and SLIC Algorithm

摘要: 本文采用超像素分割方法，针对传统的SLIC算法进行研究，通过提高分割精度来提高云检测的准确率。为了改进SLIC算法中的聚类效果，采用马氏距离代替SLIC算法中的欧氏距离进行聚类，利用阈值法对划分出的每个超像素进行二值化处理，即将云和地物分类。本文设改进前算法为O-SLIC，改进后算法为M-SLIC。在本文建立的云图数据库上将二者进行了比较，得到以下结果：M-SLIC算法的云识别率比O-SLIC算法的高；M-SLIC云检测的ROC曲线面积比O-SLIC算法大且最优分类点更靠近(0, 1)点。这些结果说明对于云图数据库中的大部分图像，M-SLIC算法相比于O-SLIC算法具有改良效果。M-SLIC算法综合了聚类和分类的特点，利用马氏距离，使算法在云检测的准确率上有所改善。

Abstract: In this paper, the superpixel segmentation method is leveraged to study the traditional SLIC algorithm, and the accuracy of cloud detection is improved by improving the segmentation accuracy. In order to improve the clustering effect in the SLIC algorithm, Mahalanobis distance is leveraged to replace the Euclidean distance in the SLIC algorithm for clustering, and the threshold method is used to binarize each divided superpixel, that is, to classify clouds and ground objects. In this paper, the algorithm before the improvement is named as O-SLIC, and the algorithm after the improvement is named as M-SLIC. The two are compared on the cloud map database established in this paper, and the following results are obtained: the cloud recognition rate of the M-SLIC algorithm is higher than that of the O-SLIC algorithm; the ROC curve area of the M-SLIC cloud detection is larger than that of the O-SLIC algorithm. The optimal classification point is closer to the (0, 1) point. These results show that for most of the images in the cloud database, the M-SLIC algorithm has an improved effect compared to the O-SLIC algorithm. The M-SLIC algorithm integrates the characteristics of clustering and classification, and leverages Mahalanobis distance to improve the accuracy of cloud detection.