基于Tetrolet变换和支持向量机的积雨云检测研究

摘要

针对卫星云图的自然纹理特点，提出了一种新的积雨云检测方法。首先利用Tetrolet变换对多种几何特征都可以实现最优逼近的特性，提取云图的频谱纹理特征，并结合传统的亮温及亮温差特征，组成特征向量集；然后通过训练支持向量机（SVM）分类器，进行积雨云检测。对FY-2D卫星云图的实验结果表明，该方法对积雨云的检测准确率达到了95%以上，相较于传统方法，具有更强的泛化能力，对雷暴等灾害天气的预警具有较高的参考价值。%According to the features of natural texture of satellite image, a novel cumulonimbus detection method was proposed. Initially, Tetrolet transform, which is able to optimally approximate multiple geometrical characteristics, was used to extract spectral texture feature of satellite image. Combined with features of traditional bright-temperature and bright-temperature difference, a feature vector set was formed. Then, by training support vector machine (SVM) classiifer, cumulonimbus detection was conducted. Experimental results of FY-2D satellite image demonstrate the accuracy of proposed method to detect cumulonimbus is more than 95%, compare with traditional methods, it has stronger generalization with higher referential value to alert disaster weather such as thunderstorm.