多结构卷积神经网络特征级联的高分影像土地利用分类

高分辨率遥感影像包含丰富的土地利用类型信息，针对单一卷积神经网络提取图像特征信息不足的问题，提出了一种多结构卷积神经网络（convolutional neural network，CNN）特征级联的分类方法。首先，选择CaffeNet（convolutional architecture for fast feature embedding）、VGG-S（visual geometry group-slow）、VGG-F（visual geometry group-fast）为实验初始模型，对网络全连接层进行参数微调，采用随机梯度下降法（stochasticgradient descent，SGD）更新网络的权重；然后以微调后的网络分别作为特征提取器对图像提取特征，级联上述3种网络的第二个全连接层输出特征作为图像表达；最后，以多类最优边界分配机（multi-class optimal margindistribution machine，mcODM）获得最终分类结果。实验采用UC Merced land-use数据集进行分类效果检验，结果表明，多结构卷积神经网络级联的方法能够达到97.55%的总体分类精度，相较于CaffeNet、VGG-S和VGG-F等，分类精度分别提升了5.71%、2.72%和5.1%。因此多结构卷积神经网络特征级联的方法能够有效提取目标特征信息，提升土地利用分类精度。

Land Use Classification Based on Multi-structure Convolution Neural Network Features Cascading

High resolution remote sensing images contain abundant information of land use types, in order to solve the problem of extracting feature information from single convolution neural network, a classification model of cascading multi-structure convolution neural networks is proposed. Firstly, We choose CaffeNet(convolutional architecture for fast feature embedding), VGG-S(visual geometry group-slow), VGG-F(visual geometry group-fast)as experimental models, and fine-tune parameter of two full-connected network layers using UC Merced land-use data set. We update the network weight by the stochastic gradient descent(SGD), and then the networks fine-tuned are used as feature extractors, the outputs of the second full-connected layer of the model above as the expression of the images. Finally, by cascading the features of the three networks, the final classification results are obtained using multi-class optimal margin distribution machine (mcODM). The experimental dataset is based on the UC Merced land-use data set. The experimental results show that the method of cascading multi-structure convolution neural network can reach 97.55%, and the accuracy of land use classification is improved by 5.71%, 2.72% and 5.1% respectively compared with pre-trained CNNs.