摘要
针对大蒜分布信息和种植面积提取准确度低等问题。该文以Sentinel-1微波和Sentinel-2多光谱遥感数据为数据源,选取大蒜返青期至鳞茎肥大期等关键物候期的极化、纹理、植被指数、红边指数、水体指数及地形等多特征进行大蒜遥感提取研究。该方法通过构建随机森林模型,利用递归特征消除算法进行特征优选,并设计4种不同特征组合的实验方案进行对比分析。结果表明,利用递归特征消除算法逐一删除评分最低特征,总体精度呈现先升高后降低的趋势,在特征数量为52时精度达到最高,总体精度是97.09%,Kappa系数是0.958 8,比特征优选前的总体精度和Kappa系数分别提高了8.37%和0.124 9,研究结果可为大区域机器学习算法提取小宗作物提供参考,对我国农业部门及时掌控大蒜种植面积提供依据。
To address the challenge of low accuracy in extracting garlic distribution information and estimating planting areas,this study utilizes Sentinel-l microwave and Sentinel-2 multispectral remote sensing data as primary data sources.It focuses on key phenological stages of garlic,from the greening stage to the bulb hypertrophy stage,and extracts multiple features such as polarization,texture,vegetation indices,red-edge indices,water indices,and topography for garlic remote sensing extraction.A random forest model was constructed,incorporating the recursive feature elimination(RFE)algorithm for feature selection.Four distinct feature combinations were designed for comparative analysis.The results demonstrate that employing the RFE algorithm to remove features with the lowest scores iteratively initially improves overall accuracy,which then decreases beyond a certain threshold.When the number of features was reduced to 52,the model achieved its highest performance,with an overall accuracy of 97.09%and a Kappa coefficient of 0.9588.This represents an 8.37%and 0.1249 improvement in overall accuracy and Kappa coefficient,respectively,compared to the model performance before feature optimization.These studies highlight the efficacy of feature optimization in enhancing the accuracy of machine learning models for extracting small-crop distribution over large areas.The research can provide a valuable reference for large-area machine learning algorithm to extract small crops and provides the basis for our agricultural department to control the garlic planting area in time.
作者
潘亚龙
余海坤
卢小平
肖丹
周俊利
PAN Yalong;YU Haikun;LU Xiaoping;XIAO Dan;ZHOU Junli(Key Laboratory of Spatio-Temporal Information and Ecological Restoration of Mines of Natural Resources of the People's Republic of China,Jiaozuo,Henan 454003,China;Henan Remote Sensing Institute,Zhengzhou 450008,China;Henan Science and Technology Innovation Center for Natural Resources,Zhengzhou 450008,China)
出处
《测绘科学》
北大核心
2025年第6期114-122,共9页
Science of Surveying and Mapping
基金
自然资源要素遥感智能识别关键技术研究及业务化应用项目(2023-382-2)
自然资源遥感智能解译样本与光谱数据库建设关键技术研究及应用示范项目(2023ZRBSHZ027)。
