利用DBN与LBP的多光谱遥感图像分类

Multispectral Remote Sensing Image Classification Using DBN and LBP

针对多光谱遥感图像分类问题，本文提出了一种利用深度置信网络（DBN）与局部二值模式（LBP）纹理特征的分类算法。首先，提取多光谱图像各波段的LBP纹理特征，并将其组合、归一化，形成144维的DBN输入特征向量；然后，计算标记好地物类别真值的像素的特征向量，将其作为训练数据对DBN网络进行训练；最后，利用训练好的网络模型完成多光谱遥感图像分类。通过天绘一号卫星多光谱遥感图像对算法进行实验验证，实验结果表明，本文算法能够取得优于神经网络（NN）和支持向量机（SVM）的性能。这说明DBN能够更好地挖掘数据的本质特征，从而提升分类的准确性。

An algorithm based on deep belief network （DBN） and local binary pattern （LBP） is proposed for multispectral remote sensing image classification. Firstly, the LBP texture features of each band in muhispeetral images are extracted, connect- ed and normalized to form a 144-dimensional eigenvector as the input value of DBN. Secondly, the eigenvectors of the pixels with true category values are calculated and marked, and the DBN network is trained based on these eigenvectors. Finally, the trained DBN network model is used to complete the muhispectral remote sensing image classification. The algorithm is verified by TH-01 satellite multispectral images, and experiment results show that the proposed algorithm can achieve better performance than neural network （NN） and support vector machine （SVM）. It proves that the DBN can accurately explore the essential characteristics of image data, thus improving the accuracy of classification.