基于卷积网络的沙漠腹地绿洲植物群落自动分类方法

为解决沙漠腹地绿洲遥感图像植物群落背景较易混淆,仅用传统的基于像元光谱信息的图像处理方法未能充分利用其图像特征信息,使得提取效果不佳的问题,针对地物类内特征复杂、类间边界模糊的特点,以连续分布的区域为研究对象,提出了一种基于深度卷积神经网络(Convolutional neural network,CNN)的高分辨率遥感影像植物群落自动分类方法。切分无人机影像获得规则块图像,利用基于CNN的VGGNet和Res Net模型分别对块图像的特征进行抽象与学习,以自动获取更加深层抽象、更具代表性的图像块深层特征,从而实现对植物群落分布区域的提取,以原图像与结果图像叠加的形式输出植物群落自动分类结果。采用了不同梯度的样本数量作为训练样本,利用文中提出的方法分析了不同梯度的训练样本数量对自动分类结果的影响。实验结果表明,训练样本数量对分类精度具有明显的影响;提高其泛化能力后,Res Net50模型与VGG19模型的建模精度从86. 00%、83. 33%分别提升到92. 56%、90. 29%; Res Net50模型分类精度为83. 53%～91. 83%,而VGG19模型分类精度为80. 97%～89. 56%,与传统的监督分类方法比较,深度卷积网络明显提高了分类精度。分类结果表明,训练样本数量不低于200时,基于CNN的Res Net50模型表现出最佳的分类结果。

Automatic Classification Method of Oasis Plant Community in Desert Hinterland Based on VGGNet and ResNet Models

In order to solve the problem of remote sensing image plant community background, only the traditional image processing method based on pixel spectral information fails to make full use of its image feature information, which makes the extraction effect poor. Aiming at the complex features of plant species and the blurring of inter class boundaries, the continuous distribution of regions was taken as the research object. A high resolution remote sensing image plant community automatic classification based on the convolutional neural network (CNN) was proposed. The UAV images were segmented to obtain regular block images, and the features of block images were abstracted and learned by CNN based VGGNet and ResNet models to automatically acquire deeper abstract and more representative image block deep features. The extraction of the plant community distribution area was performed to output the automatic classification results of the plant community in the form of superposition of the original image and the result image. The number of samples with different gradients was used as the training sample. The influence of the number of training samples with different gradients on the automatic classification results was analyzed by the proposed method. The experimental results showed that the number of training samples had a significant impact on the classification accuracy. After improving its generalization ability, the modeling accuracy of ResNet50 model and VGG19 model was improved from 86.00% and 83.33% to 92.56% and 90.29%, respectively. The classification accuracy of ResNet50 model was varied from 83.53% to 91.83%, while the classification accuracy of the VGG19 model was varied from 80.97% to 89.56%. Compared with the traditional supervised classification method, the deep convolution network significantly improved the classification accuracy. Through the analysis of classification result, it was found that the number of training samples should not be less than 200, and the CNN based ResNet50 model showed the best classification results.