| 摘要: |
| 准确识别云对提升天气预报和气候预测准确性有着重要意义，传统的阈值法和聚类法很难找出统一通用的阈值标准和方法，随着机器学习在云分类领域的应用和发展，在分类速度和分类精度上都有了明显提升。本实验对风云二号G星的红外云图进行预处理并构建卫星云图样本库，通过提取云图纹理特征再结合支持向量机（SVM）、随机森林（RF）和XGBoost分类器实现了对“晴空”、“层积云或高积云”、“积雨云”、“密层云”和“卷层云”的分类，实验结果表明：①三种分类器对该实验云分类的平均准确率分别为RF（62.5%）＞XGBoost（61.7%）＞SVM（60.0%）；②三种分类器对“层积云或高积云”的分类都最好且稳定，平均分类精度均达到了90%以上，最高为91.5%；③SVM对密层云（67.9%）、RF对卷层云（68.9%）、XGboost对晴空（68.3%）的分类效果次之，平均分类精度均达67%以上。 |
| 关键词: 灰度共生矩阵;纹理特征;卫星云图;支持向量机;随机森林 |
| DOI： |
| 投稿时间：2022-10-29修订日期：2023-08-25 |
| 基金项目:贵州省气象局研究型业务关键技术公关团队（GGTD-202212）、贵州省山地气候与资源重点实验室基金项目（QHLSSLJ[2022]-12） |
|
| Classificaiton of Satellite Cloud Image Based on Texture Features and Machine Learning |
| GU Tianhong,DU Xiaoling,LI Li,ZHU Yulei,ZHANG Yanmei,WU Changhang,LI Diannan |
| (Guizhou Meteorological Observatory) |
| Abstract: |
| Accurate identification of clouds is of great significance to improve the accuracy of weather forecast and climate prediction. It is difficult for traditional threshold method and clustering method to find a unified and universal threshold standard and method. With the application and development of machine learning in the field of cloud classification, the classification speed and classification accuracy have been significantly improved.In this experiment, the infrared cloud image of FY-2G was preprocessed and the satellite cloud image sample library was constructed.The texture features of the cloud image were extracted and combined with the SVM, RF and XGBoost classifiers to realize the classification of "clear sky", "Stratocumulus or Altocumulus", "cumulonimbus", "dense stratus" and "cirrostratus",and the experimental results show that: (1) The average accuracy of the three classifiers for the cloud classification is RF (62.5%) > XGBoost (61.7%) > SVM (60.0%).（2）The classification of “Stratocumulus or Altocumulus” is the best and the most stable, and the average classification accuracy reaches more than 90%, the highest is 91.5%.（3）SVM classification of dense stratus cloud(67.9%), RF classification of cirrostratus cloud(68.9%) and XGboost classification of clear sky（68.3%）are the next best, with an average classification accuracy of more than 67%. |
| Key words: GLCM; texture feature;satellite cloud image;SVM;RF;XGBoost |
