【摘要】 采用 GMS- 5红外 ( 1 0 .5～ 1 2 .5μm)和可见光 ( 0 .55～ 0 .9μm)两通道资料 ,采集了 1 999年 7— 1 0月中国东南沿海 57区、58区和 59区包括晴空在内的 1 2类云目标样本 2 91 2个 ,采样窗尺寸为 8× 8像素 ,随机生成训练和测试两个样本子集。对径向基函数网络 ( radial base function neural network,RBF)在云分类问题研究中的应用价值进行了全面的测试与分析 ,得到了肯定的结论 ,提出了优化设计的方法。对6类云型分类试验 ,平均正确率为 86 % ;对 1 1类云型分类试验 ,平均正确率为6 7%。采用自组织竞争神经网络实现寻找 RBF神经网络的隐层神经元中心。在特征空间生成过程中 ,采用小波包分解算法实现模式特征抽出。结果表明 ,小波包分解特征能很好地描述不同云型的差异。更多还原

【Abstract】 This paper presents an automated and efficient cloud classification scheme based on radial base function neural network (RBF),which has an average classification accuracy of more than 86% for less than 5 cloud patterns and 67% for more than 10 cloud patterns.An additional self-organized competitive neural network is also suggested to find out the center of the hidden layer neurons of RBF network,which greatly ameliorates the efficiency of the RBF classifier.Features abstracted by using the two-dimensional wavelet packet 3-level decomposition provide essentials for the description of cloud patterns,thus improving the capability of pattern recognition of neural network classifier.The input dataset employed in the scheme are defined by 2912 samples of 8×8 pixels size taken in July-October 1999 over SE China maritime regions from the visible (0.55～0.9μm) and infrared (10.5～12.5μm) channels of Geostationary Meteorological Satellite 5 (GMS-5).These images in Lambert Conformal Conic Projection and in a coarsened resolution of 13.03km×13.22km for both channels are downloaded from the meteorological operational networks. Each of the samples in the dataset is classified into one of the eleven predefined classes in accord with the SYNOP codes used in weather reports.更多还原