基于高斯混合模型的遥感影像连续型朴素贝叶斯网络分类器

提出了一种新的嵌入高斯混合模型(GMM,Gaussian Mixture Model)遥感影像朴素贝叶斯网络模型GMM-NBC(GMMbased Na ve Bayesian Classifier)。针对连续型朴素贝叶斯网络分类器中假设地物服从单一高斯分布的缺点,该方法将地物在特征空间的分布用高斯混合模型来模拟,用改进EM算法自动获取高斯混合模型的参数;高斯混合模型整体作为一个子节点嵌入朴素贝叶斯网络中,将其输出作为节点(特征)的中间类后验概率,在朴素贝叶斯网络的框架下进行融合获得最终的类后验概率。对多光谱和高光谱数据的分类实验结果表明,该方法较传统贝叶斯分类器分类效果要好,且有较强的鲁棒性。

Continuous Bayesian Network Classifier for Remote Sensing Images Based on Improved Gaussian Mixture Model

This paper proposed a new remote sensing images naive Bayesian network model GMM-NBC (GMM based na(1)ve Bayesian classifier) embedded Gaussian mixture model. To aim at the fault of single Gaussian distribution hypothesis in continuous naive Bayesian network classifier, this method simulates the distribution in the feature space using Gaussian mixture model,and gets sub-Gaussian distributions and its parameters of GMM automatically using improved EM algorithm. Gaussian mixture model as a node be embedded into naive Bayesian network, takes the output of Gaussian mixture model as the middle class posterior probability, and obtains the final posterior probability under na(1)ve Bayesian network framework. Experiments of multispectral and hyperspectral images, indicate that the performance of this method is better than traditional Bayesian network classifier, and with strong robustness.