基于卷积神经网络的沙丘形态自动分类方法  

Automatic classification of sand dune morphologyy based on convolutional neural networks

作　　者：李明[1] 燕洁华 叶汪忠 董帅 杨泽坤 LI Ming;YAN Jiehua;YE Wangzhong;DONG Shuai;YANG Zekun(Energy and Transportation Engineering College,Inner Mongolia Agricultural University,Hohhot 010018,China)

机构地区：[1]内蒙古农业大学能源与交通工程学院,呼和浩特010018

出　　处：《干旱区资源与环境》2024年第5期121-129,共9页Journal of Arid Land Resources and Environment

基　　金：国家重点研发计划项目(2018YFC0507102);内蒙古自治区直属高校基本科研业务费基础研究项目(BR220139);内蒙古自治区高等学校科学研究项目(NJZY22520)资助。

摘　　要：为探究沙丘形态数据集构建及自动分类方法,解决沙丘形态信息数据库缺失等问题,以内蒙古西部典型沙丘为研究对象,通过无人机正射影像技术采集6种典型沙丘形态数据,并结合GF-2号遥感数据采用数据增强方式构建沙丘形态数据集。通过迁移学习策略的VGGNet和ResNet模型对沙丘形态的深层语义特征进行解析与学习,自动获取更具有代表性的图像纹理特征,以此提出一种基于卷积神经网络(CNN)提取不同沙丘形态特征自动分类的方法。结果表明,基于迁移学习的VGG16模型在四种模型中分类准确率最高,达到88.14%;优化后的ResNet18模型与ResNet50模型的分类精度分别从84.04%、85.25%提升到92.79%、88.91%;优化后的ResNet18+模型表现出最佳的分类效果,准确率达到92.79%,更适用于沙丘形态的高精度自动分类。In order to address the issues of constructing a sand dune morphology dataset and developing an automatic classification method for sand dune morphology,especially in the context that there is few sand dune morphology information in existing databases,this study focuses on typical sand dunes in western Inner Mongolia to develop a method for classifying sand dunes according to their forms.Aerial images of six typical sand dune morphologies are collected using unmanned aerial vehicle orthoimage technology.Additionally,the sand dune morphology dataset is constructed by augmenting the data with GF-2 satellite remote sensing data.By utilizing the VGGNet and ResNet models with a transfer learning strategy,the deep semantic features of sand dune morphology are analyzed and learned,thus more representative texture features are automatically extracted.Based on this,a method for automatically classifying different sand dune morphological features using Convolutional Neural Networks(CNN)is proposed.The results show that the VCG16 model,based on transfer learning,achieves the highest classification accuracy among the four models,with an accuracy of 88.14%.The optimized ResNet18 and ResNet50 models improve their classification accuracies from 84.04%and 85.25%to 92.79%and 88.91%,respectively.The optimized ResNet18+model demonstrates the best performance,with an accuracy of 92.79%,making it more suitable for high-precision automated classification of sand dune morphology.

关 键 词：沙丘形态 卷积神经网络 自动分类 VGGNet ResNet 

分 类 号：P931.3[天文地球—自然地理学]