基于一维Kohonen网络的遥感影像分类方法研究

Researches on Classification Method of Multi-spectral Remote Sensing Image Based on One-dimensional Kohonen Neural Network

分析了Kohonen网络的训练模式和聚类特性,选用规模相对较小的一维Kohonen网络,并调整网络输出层的规模和邻域形状,优化网络结构;同时根据多光谱遥感影像中地物波谱曲线特征,通过不同波段组合、波段权重系数调整等方法对输入数据进行预处理,使该方法更适用于多光谱遥感影像分类和专题提取。本文以浙江省绍兴地区多光谱遥感影像分类为例,研究结果表明使用改进后的分类方法可以有效提高分类精度。

According to the training mode and clustering specialty based on Kohonen rule, One-dimensional Kohonen network with smaller size and less eomputing expenditure is devised. Combined with the characteristics revealed by the spectrum curve of various substanees, the input data should be processed in advance on multi-band combination and weight adjustment, and the neighborhood of each nerve ceil on the output layer should be modified to improve the classification and extract some geographic features. Taking Shaoxing as test area, this paper shows that modified Kohonen method could avoid the limitation of the network and improve the classification precision.