遥感结合机器学习利用优化的训练样本识别山东省灌溉农田  

Integration of remote sensing and machine learning for identifying irrigated farmland in Shandong Province of China using optimized training samples

作　　者：孙佑涛 于砚宁 葛冰洋 张佳华 白雲 吴喜芳 杨姗姗 张莎[2] SUN Youtao;YU Yanning;GE Bingyang;ZHANG Jiahua;BAI Yun;WU Xifang;YANG Shanshan;ZHANG Sha(College of Computer Science and Technology,Qingdao University,Qingdao 266071,China;College of Geographical Sciences,Hebei Normal University,Shijiazhuang 050024,China;School of Surveying and Mapping and Land Information Engineering,Henan Polytechnic University,Jiaozuo 454003,China)

机构地区：[1]青岛大学计算机科学技术学院,青岛266071 [2]河北师范大学地理科学学院,石家庄050024 [3]河南理工大学测绘与国土信息工程学院,焦作454003

出　　处：《农业工程学报》2025年第6期154-164,共11页Transactions of the Chinese Society of Agricultural Engineering

基　　金：国家自然科学基金项目(42201407,42101382);山东省自然科学基金项目(ZR2022QD120);河南理工大学测绘科学与技术“双一流”学科创建项目(GCCYJ202427)。

摘　　要：了解灌溉农田分布对合理利用水资源、及时调整农业生产政策和保障粮食安全具有重要意义。已有基于机器学习方识别灌溉农田的研究多采用二值化样本标注(即仅标注为灌溉和非灌溉),可能会导致灌溉样本遗漏,造成灌溉农田的识别精度较低。为避免该问题,该研究提出一种为训练样本赋值灌溉分数的方案。首先,使用归一化植被指数、增强型植被指数和绿度植被指数结合统计灌溉面积数据分别生成3幅初步灌溉地图;然后,结合3幅不同的初步灌溉地图,使用该研究提出的方案为每个像元赋值不同的灌溉分数并获取训练样本;再使用随机森林(random forest,RF)和卷积神经网络(convolutional neural network,CNN)2种机器学习方法分别预测研究区逐年逐像元的灌溉分数,识别山东省2018—2022年250 m分辨率的灌溉农田;最后,通过验证样本和统计数据进行验证,获取山东省2018—2022年的灌溉农田空间分布,并分析灌溉农田的时空分布特征。结果表明:1)相比2种二值化处理获取训练样本的方法,使用该研究提出的方案获取训练样本并识别灌溉农田时,RF和CNN识别灌溉农田的县级R~2高达0.95和0.93;RF识别山东省灌溉农田的精度评价指标均优于CNN,且在不同年份表现较为稳定;2)2018—2022年山东省的灌溉农田空间分布较为一致,主要分布在鲁西北及鲁南地区,胶东半岛及鲁中地区分布较为稀少,近5年山东省的统计灌溉面积呈小幅增长趋势,2022年识别灌溉农田面积相较于2018年也有小幅增长。该研究提出的针对训练样本进行赋值灌溉分数的方案能够有效提高山东省灌溉农田识别的精度,是识别区域尺度灌溉农田的一种可靠、有效的方法。Understanding the distribution of irrigated farmland is of great significance for rational utilization of water resources,timely adjustment of agricultural production policies and protection of food security.Existing studies based on machine learning to identify irrigated farmland mostly use binarized sample labeling(that is,only labeling irrigated and nonirrigated),which may lead to the omission of irrigation samples,resulting in low identification accuracy of irrigated farmland.To avoid this problem,a scheme of assigning irrigation fractions to training samples was proposed in this paper.Firstly,three preliminary irrigation maps were generated using normalized vegetation index,enhanced vegetation index and greenness vegetation index combined with statistical irrigation area data.Then,combined with three different preliminary irrigation maps,the scheme proposed in this study assigned different irrigation scores to each pixel and obtained training samples.Then,two machine learning methods,random forest(RF)and convolutional neural network(CNN),were used to predict the pixel-bypixel irrigation fraction of the study area,respectively.Identification of 250 m resolution irrigated farmland in Shandong Province from 2018 to 2022 was conducted.Finally,the spatial distribution of irrigated farmland in Shandong Province from 2018 to 2022 was obtained based on verification samples and statistical data,and the spatial and temporal distribution characteristics of irrigated farmland were analyzed.The results showed as follows:1)Compared with the two binarization methods for obtaining training samples,the county R2 of RF and CNN for identifying irrigated farmland was as high as 0.95 and 0.93 when the proposed scheme was used to obtain training samples and identify irrigated farmland;The accuracy of RF identification of irrigated farmland in Shandong Province was better than that of CNN,and the performance was relatively stable in different years.2)From 2018 to 2022,the spatial distribution of irrigated farmland in Shandong Prov

关 键 词：机器学习 随机森林 卷积神经网络 训练样本赋值 灌溉农田识别 

分 类 号：S127[农业科学—农业基础科学]