夜间静止卫星红外云图的GHSOM网络云分类模型

摘要

针对夜间云分类准确率低下的问题，利用奇异值分解方法对FY-2E夜间红外云图进行特征提取和选择，从中筛选出包括亮温和分裂窗差值在内的不同的纹理特征。分别采用动态增长型分层自组织和自组织映射2种神经网络模型对夜间云图进行分类，并将2种网络模型的分类效果进行对比分析。实验结果表明：GHSOM网络模型在夜间云图分类方面效果较好，平均准确率总体上高于SOM，通过分层的分类方法极大地提高了夜间云图的分类准确率。%Aiming at the low accuracy of cloud classification at night,the features of FY-2E cloud images which in-clude bright temperatures and split window values were extracted and selected based on the method of singular value decomposition. The neural network models of growing hierarchical self-organizing map( GHSOM)and self-organi-zing map( SOM)were built separately to classify cloud images at night,meanwhile,contrasting the classified effect of the two network models. The experiments results showed that GHSOM network can improve the distinguishing effect of cloud images at night greatly through hierarchical classified method,and the average accuracy of cloud clas-sification results is higher than SOM.