首页|基于改进密集连接网络的土地卫片场景分类方法

基于改进密集连接网络的土地卫片场景分类方法

扫码查看
原文链接
  • 国家科技期刊平台
  • NETL
  • NSTL
  • 万方数据
为提高土地卫片图像场景识别的效率和精度,本文构建拉普拉斯金字塔-密集连接卷积网络(Lap-DenseNet)模型对土地场景应用进行识别分类,在Lap-DenseNet模型中Lap采用三层金字塔结构,DenseNet模型选用169层结构.将构建好的Lap-DenseNet模型应用到含有6种土地场景的卫片分类中,结果显示:Lap-DenseNet模型训练集的迭代次数不宜过多,否则会因为过拟合现象导致分类效果降低,当迭代次数为200次时分类效果最佳;Lap-DenseNet模型对农村道路分类效果最好,对以绿色背景为主的耕地复耕、未建设用地、农用地复绿分类效果较差,6种场景的平均分类准确率为93.66%;与谷歌卷积网络(GoogLeNet)、快速特征嵌入卷积网络(CaffeNet)、基于密集连接的双流深度特征融合卷积网络(TEX-TS-Net)、基于VGG16的附加资源卷积网络(ARCNet-VGG16)、基于Inception-v3的胶囊卷积网络(Inception-v3-CapsNet)、基于全局上下文空间注意和密集连接的卷积网络(GCSANet)共6种场景分类方法相比,Lap-DenseNet模型的分类效果最好,可在土地卫片场景分类工作中予以合理运用.
Satellite image classification method for land scenes based on improved densely connected network
In order to improve the efficiency and accuracy of satellite image identification of land scenes,a Laplace pyramid-densely connected convolutional network(Lap-DenseNet)model was constructed to identify and classify land scenes.Lap in the Lap-DenseNet model adopted a three-layer pyramid structure,while DenseNet used a 169-layer structure.The constructed Lap-DenseNet model was applied to satellite image classification containing six types of land scenes.The results show that the number of iterations in the training set of the Lap DenseNet model should not be too much.Otherwise,the classification performance will be reduced due to overfitting.The best classification performance is achieved when the number of iterations is 200.The Lap DenseNet model has the best classification performance for rural roads but has poor classification performance for cultivated land with a green background,uncultivated land,and agricultural land with a green background.The average classification accuracy for the six types of land scenes is 93.66%.Compared with six scene classification methods such as Google convolutional network(GoogLeNet),convolutional network for fast feature embedding(CaffeNet),dense connectivity-based two-stream deep feature fusion convolutional networks(TEX-TS-Net),convolutional network of additional resources based on VGG16(ARCNet-VGG16),capsule convolutional networks based on Inception-v3(Inception-v3-CapsNet),and convolutional networks based on global context space attention and dense connections(GCSANet),the Lap DenseNet model has the best classification performance and can be reasonably applied in satellite image classification work of land scenes.

satellite images of land scenesscene classificationLaplace pyramid-densely connected convolutional network(Lap-DenseNet)modelnumber of iterationsclassification accuracy

吴志斌

展开 >

海南国源土地矿产勘测规划设计院,海南 海口 570203

土地卫片 场景分类 拉普拉斯金字塔-密集连接卷积网络(Lap-DenseNet)模型 迭代次数 分类准确率

北京市自然科学基金

4214069

2024

10.19580/j.cnki.1007-3000.2024.09.018

北京测绘
北京市测绘设计研究院,北京测绘学会

北京测绘

影响因子：0.55
ISSN：1007-3000
年,卷(期)：2024.38(9)