摘要
在遥感图像场景分类中,基于卷积神经网络(convolutional neural network,CNN)的分类算法存在对训练数据的依赖性,且在缺乏训练数据时分类效果差等问题,提出一种基于迁移学习的分类算法。首先,选取现有的多个CNN预训练模型,利用迁移学习的优势对模型进行微调,目的是提取图像不同的高层特征;然后,融合图像的多种高层特征,使得特征信息更加丰富;最后,将融合后的高层特征输入到基于逻辑回归的遥感图像分类器中,得到遥感影像的分类结果。在UCMerced_LandUse遥感数据集中进行实验,与现有算法进行比较分析,所提算法在3种评价指标上有明显提升。通过分析实验结果表明,该算法在仅有10%的训练数据下,能够达到92.01%的分类准确率和91.61%的Kappa系数。
In remote sensing image scene classification,a classification algorithm based on convolutional neural network(CNN)has the dependence on training data,and the classification effect is poor in the absence of training data,and a classification algorithm based on transfer learning is proposed.Firstly,the existing pre-training model of multiple CNN is selected,and the model is fine-tuned by using the advantages of transfer learning to extract the different high-level features of the image,then,the fusion of the image′s many high-level features makes the feature information more abundant,and finally,the merged high-level features are input into the remote sensing image classifier based on logical regression,and the classification results of remote sensing images are obtained.Experiments are carried out in remote sensing data sets of UCMerced_LandUse,and the existing algorithms are compared and analyzed,and the proposed algorithms are significantly improved in three evaluation indicators.By analyzing the experimental results,it is shown that the algorithm can achieve 92.01%classification accuracy and 91.61%Kappa coefficient under only 10%of the training data.
作者
刘有耀
陈琪
李舒曼
LIU Youyao;CHEN Qi;LI Shuman(School of Electronics and Engineering,Xi′an University of Posts and Telecommunications,Xi′an,Shaanxi 710121,China)
出处
《光电子．激光》
CAS
CSCD
北大核心
2022年第7期709-714,共6页
Journal of Optoelectronics·Laser
基金
国家自然科学基金(61874087,61834005,61634004)资助项目
关键词
遥感影像
场景分类
卷积神经网络
迁移学习
逻辑回归
remote sensing image
scene classification
convolutional neural network(CNN)
transfer learning
logical regression
