为了减少高光谱图像中的冗余以及进一步挖掘潜在的分类信息，本文提出了一种基于特征重要性的卷积神经网络（ convolutional neural networks，CNN）分类模型。首先，利用贝叶斯优化训练得到的随机森林模型（ random forest，RF）对高光谱遥感图像进行特征重要性评估；其次，依据评估结果选择合适数目的高光谱图像波段，以作为新的训练样本；最后，利用三维卷积神经网络对所得样本进行特征提取并分类。基于两个实测的高光谱遥感图像数据，实验结果均表明：相比原始光谱信息直接采用支持向量机（ support vector machine，SVM）和卷积神经网络的分类效果，本文所提基于特征重要性的高光谱分类模型能够在降维的同时有效提高高光谱图像的分类精度。

To reduce the redundancy in hyperspectral images and further explore their potential classification information, a convolutional neural network(CNN) classification model based on feature importance is proposed. First, the random forest(RF) model obtained by Bayesian optimization training is used to evaluate the importance of hyperspectral images. Second, an appropriate number of hyperspectral image bands are selected as new training samples according to the evaluation results. Finally, the 3D-CNN is used to extract and classify the obtained samples. Based on two sets of measured hyperspectral remote sensing image data, the experimental results demonstrate the following: compared with the original spectral information obtained directly using a support vector machine(SVM) and the CNN classification effect, the proposed hyperspectral classification model based on feature importance can effectively improve the classification accuracy of hyperspectral images while reducing dimensionality.（cstc2014yykfB30003）。