摘要
在对湿地遥感信息提取的过程中,当特征分析深度达到一定水平时,便会出现退化问题,导致提取结果存在较大误差,因此基于深度学习研究了湿地遥感信息的提取方法。首先,对遥感信息进行预处理,以位置固定作为控制点选择的标准。将选定的控制点均匀分布在遥感图影像中,再利用二元多项式原理对原始的遥感图像像素坐标进行变换处理。其次,采用深度学习中的深度卷积神经网络,提取经过预处理的遥感图像信息,并在网络中引入残差块,以避免退化问题的出现。最后,个性化设置ReLU激活函数,实现对不同信息的针对性提取,以减小提取结果误差。测试结果表明,设计方法对于湿地温度植被干旱指数的提取结果与实际参数之间的误差稳定在0.01内,明显低于对比方法,说明设计方法有一定的实际应用意义。
In the process of remote sensing information extraction for wetlands,when the depth of feature analysis reaches a certain level,the degradation problem will occur,leading to large errors in the extraction results.Therefore,this paper studies the extraction method of wetland remote sensing information based on deep learning.Firstly,the remote sensing information is preprocessed and the fixed position is used as the standard for selecting control points.After the selected control points are evenly distributed in the remote sensing image,the pixel coordinates of the original remote sensing image are transformed by using the principle of binary polynomial.Secondly,the deep convolutional neural network in deep learning is used to extract information from the preprocessed remote sensing image,and residual blocks are introduced into the network to avoid the occurrence of degradation problem.Finally,the ReLU activation function is personalized set to achieve targeted extraction of different information,so as to reduce the error of extraction results.The test results show that the error between the extraction results of temperature vegetation drought index and the actual parameters of the design method is stable within 0.01,which is obviously lower than that of the comparison method,indicating that the design method has certain practical application significance.
作者
张文博
张源
ZHANG Wenbo;ZHANG Yuan(Changsha Environmental Protection College,Changsha Hunan 410000,China;Hunan Polytechnic of Water Resources and Electric Power,Changsha Hunan 410000,China)
出处
《信息与电脑》
2022年第23期173-175,共3页
Information & Computer
基金
2021年度湖南省教育厅科学研究项目(项目编号:21C1586)。
关键词
深度学习
湿地遥感
信息提取
二元多项式原理
深度卷积神经网络
残差块
deep learning
wetland remote sensing
information extraction
principle of binary polynomials
deep convolutional neural networks
residual block
