基于Kohonen神经网络的遥感影像监督分类

Research on remote sensing image classification based on Kohonen neural network

作为一种自组织网络,Kohonen神经网络在遥感影像聚类中得到了广泛的应用。但Kohonen算法具有初始权值随机化、学习率和邻域难以确定等缺陷,并且作为一种聚类算法,难以直接进行监督分类。对Kohonen非监督学习算法进行改进,用最大最小距离法确定初始权值,利用权值误差作为网络训练收敛条件,并将Kohonen算法与奖惩学习结合起来,进行遥感监督分类。最后,选取安徽省寿县附近区域作为实验区,利用环境一号卫星影像(HJ-1B/CCD)进行了分类实验,其总精度为89.48%,Kappa系数为0.87。实验结果表明,改进的Kohonen神经网络遥感监督分类算法,其分类精度优于经典的最大似然法,取得了较为理想的分类结果。

Kohonen algorithm has defects that initial weights are randomized, learning rate and neighborhood are ditticult to determine, and as a clustering algorithm, kohonen algorithm is difficult to be used directly for supervised classification. In this paper, we used max-rain distance method to determine Kohonen Network＇ initial weights, introduced weight error as the convergence condition of network, and improved Kohonen network＇ unsupervised-learning algorithm. We then used the im- proved Kohonen＇ learning algorithm combined with penalty formula to carry out supervised elassitication of remote sensing data. Experiments showed that higher accuracy resulted from this improved algorithm.