基于ResNet18和随机森林的遥感图像复杂场景分类方法
A Complex Scene Classification Method of Remote Sensing Images Based on ResNet18 and Random Forest
彭程 1王莉 1王安邦 1齐涛 1王慧 2王靖伟1
作者信息
- 1. 日照市自然资源和规划局,山东日照 276800
- 2. 日照市岚山区发展和改革局,山东日照 276800
- 折叠
摘要
复杂场景分类是遥感图像解译的一项重要内容.本文通过优化ResNet18深度残差网络和随机森林,实现了遥感图像复杂场景的高精度分类.首先通过数据扩充将数据库扩充以缓解因训练样本少带来的过拟合问题,然后采用ResNet18深度残差网络自动提取遥感图像场景特征,最后使用随机森林分类器实现复杂场景分类任务并分别在NWPU-RESISC45和UC Merced Land Use数据库上进行了实验.结果表明,本文模型场景分类准确率分别为98.86%和99.17%,与单独使用ResNet18深度残差网络相比,本文模型分类准确率分别提高3.36%和1.71%,相比于其他场景分类方法,本文模型分类准确率分别提高5.23%和1.55%.
Abstract
Complex scene classification is a crucial aspect of remote sensing image interpretation.This paper achieves high-precision classification of complex scenes in remote sensing images by optimizing the ResNet18 deep residual network and Random Forest.First,data augmentation is used to expand the database,alleviating the overfitting problem caused by the limited number of training samples.Then,the ResNet18 deep residual network is employed to automatically extract scene features from the remote sensing images.Finally,a Random Forest classifier is used to accomplish the complex scene classification task.Experiments were conducted on the NWPU-RESISC45 and UC Merced Land Use databases.The results show that the scene classification accuracies of the proposed model are 98.86%and 99.17%,respectively.Compared to using the ResNet18 deep residual network alone,the proposed model improves classification accuracy by 3.36%and 1.71%,respectively.Moreover,in comparison with other scene classification methods,the proposed model improves classification accuracy by 5.23%and 1.55%,respectively.
关键词
数据扩充/深度残差网络/随机森林/遥感图像/场景分类Key words
Data augmentation/deep residual network(ResNet)/Random Forest/remote sensing images/scene classification引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据