改进AlexNet的高分三号全极化海岛地物分类

Island’s Ground Object Classification of Full Polarimetric GF-3 Imagery Using Improved AlexNet

针对经典卷积神经网络难以有效分类全极化SAR数据中复杂的海岛地物的问题,为满足日益精细化的监测需求、充分发挥SAR在海岛监测中的作用,文章对经典的AlexNet改进,提出了一种应用于全极化SAR数据海岛地物分类的卷积神经网络模型。该模型是在AlexNet基础上调整卷积核大小及全连接层,减少参数,加入池化层,降低维度,减少计算复杂度。利用高分三号卫星对南日岛进行观测获取的全极化SAR图像进行实验和分析,表明该方法能够对全极化SAR图像中海岛的多类地物进行有效区分,与AlexNet的分类结果相比,精度提升5.56%。

In view of the problem that the classical convolution neural network is difficult to effectively classify the complex island’s ground objects in the full polarimetric SAR data,in order to meet the increasingly refined monitoring needs and give full play to the role of SAR in island monitoring,this paper proposes a convolutional neural network model for island’s ground object classification based on full polarimetric SAR data,which is improved by AlexNet.The model adjusts the size of convolution kernel and full connection layer on the basis of AlexNet to reduce parameters,and adds pooling layer to reduce dimension and computational complexity.The experiment and analysis of the polarimetric SAR images obtained from the observation of Nanri island by GF-3 satellite show that this method can effectively distinguish the multi types of island’s ground objects in the polarimetric SAR images.Compared with the classification results of AlexNet,the accuracy is improved by 5.56%.