多光谱卫星云图的SOFM-PNN网络耦合的云分类模型

A Cloud Classification Model of Multi-spectrum Satellite Cloud Images Based on the Network Coupling SOFM with PNN

针对单一类型的神经网络分类器难以正确区分和有效识别复杂云类特征的缺陷,本文基于静止气象卫星云图多光谱云类样本,通过计算、分析云图灰度、梯度与纹理特征,提取了云分类最佳判别因子,建立了自组织网络(SOFM)与概率神经网络(PNN)的综合云分类器优化模型.该分类器首先利用自组织网络对云类样本进行无监督初始分类,确定出相似样本子集;随后用概率神经网络对初始分类误差进行有监督修正和分类结果的二次优化判别.试验结果表明,该分类器可有效提高云类判别效果,分类结果的总正确率达到92.4%,Kappa系数为90.82,明显优于单一的统计分类器判别效果.

Based on the multi-spectrum samples of stationary meteorology satellite cloud images, the best distinguishing factors of cloud classification were distilled and a synthetical optimized cloud classifier combining the advantages of both self-organizing feature map （SOFM） and probabilistic neural network （PNN） was established by computing and analyzing the gray- gradient co-occurrence matrix and texture characters of satellite cloud image samples to overcome the shortcoming of single a neural network （ANN） classifier difficult to identify and classify complex cloud characters accurately and effectively. Firstly, the cloud samples were classified to identify and partition the analogical sample-sets by SOM without supervision, then the initial classification error was revised under the supervision and the initial classification results were optimized once more by using PNN. The experiments results showed that the synthetical SOM-PNN classifier can improve the distinguishing effect in that the total accuracy of cloud classification results could reach 92. 4% and the coefficient of Kappa was 90. 82, which excel other single-statistical classifier evidently.