Hyperspectral Image Classification Based on Adaptive Ridgelet Neural Network
基于自适应脊波网络的高光谱遥感图像分类

Artificial neural network is an important tool in the scope of remote sense classification. A new model named adaptive ridgclet 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 fcaturcselected 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 RI3F 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.