基于边缘特征增强的油菜田块信息提取方法  

Edge Feature Enhancement-based Field Extraction Method for Oilseed Rape

作　　者：张宁 孔开昕 许卫 董航 段金馈 张承明[1] ZHANG Ning;KONG Kai-xin;XU Wei;DONG Hang;DUAN Jin-kui;ZHANG Cheng-ming(College of Information Science and Engineering/Shandong Agricultural University,Tai'an 271018,China;Shandong Provincial Climate Centre,Ji'nan 250031,China)

机构地区：[1]山东农业大学信息科学与工程学院,山东泰安271018 [2]山东省气候中心,山东济南250031

出　　处：《山东农业大学学报(自然科学版)》2025年第2期203-213,共11页Journal of Shandong Agricultural University:Natural Science Edition

基　　金：青海省科技计划项目“基于遥感图像超分辨率技术的油菜地土壤水分监测”(2021-ZJ-739)。

摘　　要：获取高质量的特征是从遥感影像中提取高精度油菜田块信息的关键步骤,本文针对油菜田块边缘像素特征质量通常不够理想的问题,提出了一种边缘特征增强的油菜田块信息提取模型(Edge Feature Enhancement Net,EFENet)。首先,采用编码器-解码器结构作为特征提取器基本结构,设计了边缘注意力机制(Edge Attention Mechanism, EAM)作为特征增强模块,由通道注意力和空间注意力两个子模块组成,用于提高边缘像素的特征质量。其次,设计了考虑边缘增强的损失函数(Considering Boundary Enhancement Loss Function, CBELoss)以提高边缘特征质量,由权重项ω和特征质量数学期望(Mathematical Expectation of Characteristic Mass, MECM)组成,ω根据像素空间相关性来表达样本位置对误差的影响;MECM通过评价样本质量以增强特征的区分度。EFENet采用SoftMax作为逐像素分类器。选择GF-6 PMS (Gaofen-6 Panchromatic and Multispectral Scanner)影像为数据源;青海省海北藏族自治州门源回族自治县为研究区;ERFNet、RefineNet、UNet为对比模型开展对比实验。结果表明,EFENet在F1分数(92.40%)、召回率(93.64%)、查准率(92.83%)和准确率(92.51%)方面均优于对比模型,表明该模型在提取油菜田块信息方面具有明显优势。Obtaining high-quality features is a key step in extracting high-precision oilseed rape plot information from remote sensing images.To address the problem of the unsatisfactory quality of the edge features of oilseed rape fields,this paper proposes an Edge Feature Enhancement Net(EFENet)for oilseed rape field information extraction model.Firstly,the encoder-decoder structure is adopted as the basic structure of the feature extractor,and the Edge Attention Mechanism(EAM)is designed as the feature enhancement module,comprising two sub-modules:the channel attention and spatial attention,to improve the feature quality of the edge pixels.Secondly,the Considering Boundary Enhancement Loss Function(CBELoss)is designed to improve the edge feature quality,consisting of a weight termωand Mathematical Expectation of Characteristic Mass(MECM).ωexpresses the effect of sample location on the error based on the pixel spatial correlation,and the MECM enhanced the discriminative degree of features by evaluating the quality of the samples.EFENet employs SoftMax as a pixel by pixel classifier.In this paper,PMS(Gaofen-6 Panchromatic and Multispectral Scanner)images are selected as the data sources,Menyuan Hui Autonomous County,Haibei Tibetan Autonomous Prefecture and Qinghai Province as the study area,ERFNet,RefineNet,and UNet as the comparison models.The results show that the proposed method outperforms the comparison model in terms of F1 score(92.40%),recall rate(93.64%),checking accuracy(92.83%),and precision(92.51%),indicating that the model has obvious advantages in extracting the information of GF-6 PMS oilseed rape fields.

关 键 词：卷积神经网络 油菜田块信息 边缘特征增强 损失函数 特征质量 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]