基于自适应脊波网络的高光谱遥感图像分类  

Hyperspectral Image Classification Based on Adaptive Ridgelet Neural Network

作　　者：孙锋利[1,2] 何明一[1] 高全华[2] 

机构地区：[1]西北工业大学电子信息学院,西安710077 [2]长安大学理学院,西安710064

出　　处：《计算机科学》2011年第8期260-264,共5页Computer Science

基　　金：国家自然科学基金(60736007);长安大学中央高校专项科研基金(CHD2010JC133)资助

摘　　要：神经网络是遥感地物自动分类的重要工具之一。利用多尺度几何分析中的脊波基函数建立了一种自适应脊波网络模型。在传统自适应粒子群算法的基础上,提出一种引入粒子密度因子的自适应粒子群优化算法作为网络训练算法。为验证其性能,利用互信息约简技术对220波段AVIRIS 92AV3C高光谱数据进行约简,并将它们作为网络输入实现对高光谱遥感地物的自动分类。仿真试验表明:引入粒子密度因子的粒子群算法与传统粒子群算法相比,不易出现早熟问题,在处理高维非线性组合优化问题时具有一定优势;由于脊波函数对高维奇异性的表征能力,相比于传统的RBF和SVM分类器,脊波神经网络分类器对具有明显边界特征的地物分类问题具有较高的精度,同时网络规模小,结构简单。Artificial neural network is an important tool in the scope of remote sense classification.A new model named adaptive ridgelet neural network was presented based on the theory of multi-scaled geometric analysis.On the bases of conventional ones,a novel adaptive PSO algorithm progressed by a so-called swarm density factor was proposed to train the ridgelet network constructed.To validate the performance of ridgelet network,a hyperspectral image classification task was carried out on the feature-selected hyperspectral data set AVIRIS 92AV3C by means of mutual information band-selection method.Numerical experiments show the novel PSO algorithm outperforms the conventional PSO for ridgelet network training especially in high-dimensional scenarios.Ridgelet neural network,compared with RBF and SVM classifier,is advantageous in accuracy referring to ground materials classification with apparent margin,and under the same circumstances,the network always works with simpler structure and smaller size.

关 键 词：脊波 神经网络 粒子群优化 高光谱分类 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]