基于改进VGG16网络的机载高光谱针叶树种分类研究

Study on Hyperspectral Conifer Species Classification Based on Improved VGG16 Network

为解决针叶树种识别上存在的分类精度较低和训练时间较长的难点问题。本文提出一种基于卷积神经网络的机载高光谱影像针叶树种分类的网络模型。实验选定VGG16作为基础网络进行改进,精简了网络层的结构,重新组织了卷积核的排列,更好地适应高光谱分类任务,对实验选择的茶壶实验森林的机载高光谱影像数据带进行数据增强,使用Adam优化器进行训练优化,使用学习率反向时间衰减器和Early-stooping优化器加快网络拟合的速度、增加网络的泛化能力。研究结果表明,在类间差距小、类内差距大的情况下,与对比实验效果最好的未改进VGG16网络相比,针叶树种高光谱影像多标签分类的精度提高了8.17%,达到了94.47%的分类准确率,且训练时长缩减了6倍多。从而得到:将卷积核数量按照从大到小的方式排列有助于高光谱信息的提取和训练时间的缩短;网络层数的精简可以在保证模型达到拟合的前提下训练不会过度,减小训练时间;数据增强对针叶树种识别准确率的提升有着很大的帮助。

In order to solve the difficult problems of low classification accuracy and long training time in conifer species recognition,this paper presents a coniferous species classification network model based on convolutional neural network in airborne hyperspectral images.VGG16 was selected as the basic network to improve the experiment,which simplified the structure of the network layer,reorganized the arrangement of convolution kernel,and better adapted to the task of hyperspectral classification.The airborne hyperspectral image data band of the teapot experimental forest selected for the experiment was enhanced,the Adam optimizer was used for training optimization,and the learning rate inverse time attenuator and the Early-stooping optimizer were used to accelerate the speed of network fitting and increase the generalization ability of the network.The results showed that,in the case of small inter-class gap and large intra-class gap,compared with the unimproved VGG16 network with the best comparative experiment effect,the accuracy of multi-label classification of hyperspectral images of coniferous species was improved by 8.17%,and the classification accuracy was 94.47%,and the training time was reduced by more than 6 times.It can be concluded that the arrangement of the number of convolution kernels from large to small was helpful to the extraction of hyperspectral information and shorten the training time.The simplification of network layers can ensure that the model can be fitted without excessive training and reduce the training time.Data enhancement was of great help to the improvement of coniferous species identification accuracy.