引用本文:[点击复制]
[点击复制]
【打印本页】【下载PDF全文】 查看/发表评论下载PDF阅读器关闭

←前一篇|后一篇→

过刊浏览    高级检索

本文已被:浏览次   下载 本文二维码信息
码上扫一扫!
基于改进BP神经网络的多分辨率遥感图像分类及对比分析
戚王月,胡宏祥,夏萍,周婷
0
(安徽农业大学工学院,合肥 230036;安徽农业大学资源与环境学院,合肥 230036)
摘要:
在遥感图像分类的研究中,传统的分类方法对“同物异谱”、“异物同谱”现象识别能力较差。此外,常用的BP神经网络分类存在时间长、易陷入局部极小等不足。将BP网络中的激励函数添加偏置参数、学习率进行自适应调整,并与最大似然、BP神经网络分类比较,结果表明改进的BP神经网络分类精度为89.69%,比最大似然提高了15.35%,比BP神经网络提高了23.81%。另一方面,基于改进的BP神经网络分类,对分辨率为16 m的高分一号卫星(GF-1)图像和分辨率为5.8 m的资源三号卫星(ZY-3)图像进行分类比较,并以ZY-3分类图作为检验图像,GF-1图像的分类精度达到了88.02%,各类地物的用户精度和制图精度在70%~99%之间,说明成本较低、宽幅较广的GF-1图像在地物信息获取方面可基本实现ZY-3卫星图像效果,为遥感图像地物信息提取提供了一定的参考。
关键词:  BP神经网络  遥感图像分类  高分卫星  资源卫星  多分辨率遥感
DOI:10.13610/j.cnki.1672-352x.20191013.015
基金项目:安徽省高校优秀青年人才支持计划重点项目(gxyqZD2017019), 安徽省国际科技合作计划项目(1604b0602029), 安徽省自然科学基金(1808085ME158)和安徽省高等学校自然科学研究项目(KJ2017A134)。
Multi-resolution remote sensing images classification and comparison analysis based on improved BP neural network
QI Wangyue,HU Hongxiang,XIA Ping,ZHOU Ting
(School of Engineering, Anhui Agricultural University, Hefei 230036;School of Resources and Environment, Anhui Agricultural University, Hefei 230036)
Abstract:
In the research of remote sensing image classification, traditional classification method has poor performance in recognizing “same object with different spectra” and “different objects with same spectrum” phenomenon. Moreover, traditional BP neural network classification method is time consuming and easy to fall into local minimum. In this paper, the excitation parameters in the BP network were adaptively adjusted by adding bias parameters and learning rates, and compared with the maximum likelihood neural network and BP neural network classification. Results showed that the improved BP neural network classification accuracy is 89.69%, which is 15.35% higher than the maximum likelihood neural network result, and 23.81% higher than BP neural network result. Besides, based on the improved BP neural network classification method, the high-resolution satellite (GF-1) image with resolution of 16m and the resource third satellite (ZY-3) with resolution of 5.8 m were classified and compared. Using the ZY-3 classification map as inspection, the classification accuracy of the GF-1 image reaches 88.02%, user accuracy and prod accuracy of various types of features are between 70% and 99%, indicating that the GF-1 image with lower cost and wider width can basically realize the ZY-3 satellite classification performance in terms of ground object information acquisition. The image effect provides reference for remote sensing image feature information extraction.
Key words:  BP neural network  remote sensing image classification  GF-1 satellite  ZY-3 satellite  multi- resolution remote sensing

用微信扫一扫

用微信扫一扫