基于多光谱影像的阿拉尔垦区棉田土壤盐分反演  

Monitoring soil salinization of cotton fields in the Aral Reclamation Area using multispectral imaging

作　　者：洪国军 谢俊博 张灵 范振岐[2,3] 喻彩丽 付仙兵 李旭[2,3] HONG Guojun;XIE Junbo;ZHANG Ling;FAN Zhenqi;YU Caili;FU Xianbing;LI Xu(Jiangxi University of Science and Technology Jiangxi Provincial Institute of Regional Development,Nanchang 330200,Jiangxi,China;School of Information Engineering,Tarim University,Alar 843300,Xinjiang,China;Key Laboratory of Tarim Oasis Agricultural Ministry of Education,Tarim University,Alar 843300,Xinjiang,China;College of Water Resources and Architectural Engineering,Tarim University,Alar 843300,Xinjiang,China;College of Ocean of Shanwei Institute of Technology,Shanwei 516600,Guangdong,China)

机构地区：[1]江西科技学院江西省区域发展研究院,江西南昌330200 [2]塔里木大学信息工程学院,新疆阿拉尔843300 [3]塔里木大学塔里木绿洲农业教育部重点实验室,新疆阿拉尔843300 [4]塔里木大学水利与建筑工程学院,新疆阿拉尔843300 [5]汕尾职业技术学院海洋学院,广东汕尾516600

出　　处：《干旱区研究》2024年第5期894-904,共11页Arid Zone Research

基　　金：国家自然基金地区基金(61662064);国家自然基金地区基金(42061046)资助。

摘　　要：针对新疆土壤盐分信息获取困难,无法快速、准确的大范围评估土壤盐渍化情况,本研究以新疆阿拉尔垦区的棉田为研究对象,利用Sentinel-2 SR和Landsat-9 OLI的多光谱遥感影像数据,采用穷举特征组合和交叉验证方法,从20个光谱指数和组合光谱指数构建的高维数据集中筛选出最优特征子集,并比较四种机器学习模型(即XGBoost、随机森林、深度神经网络和K-近邻)在不同特征组合下的土壤盐渍化反演精度,同时分析Sentinel-2 SR和Landsat-9 OLI遥感影像在土壤盐渍化反演中的精度差异。研究结果表明:(1)基于XGBoost算法构建的模型能够实现棉田盐渍化高精度预测,不同特征组合的R^(2)均高于0.74,MSE均低于0.04,MAPE低于0.13。(2)在特征组合1条件下,Sentinel-2 SR(S3+GBNDVI)与Landsat-9 OLI(SI+NDVI)遥感影像使用XGBoost算法均获得了最高预测精度。(3)Sentinel-2 SR影像数据在棉田盐渍化预测中的精度(R^(2)=0.73~0.88)优于Landsat-9 OLI影像数据。本研究实现了新疆阿拉尔垦区棉田土壤盐渍化精准监测,为垦区棉田土壤盐渍化治理和防治提供有效的技术参考。Given the difficulties in the field measurement of soil salinization in Xinjiang and the difficulty in quickly and broadly evaluating the potential hazards of soil salinization,this study considers cotton fields in the Aral Reclamation Area of Xinjiang as the research object,and uses multispectral remote sensing image data from Sentinel-2 SR and Landsat-9 OLI to construct a high-dimensional data set by comprehensively integrating 20 spectral indices and combining spectral indices.The optimal feature subset is screened using the method of exhaustive feature combination and cross-validation,and the inversion accuracy of soil salinization is compared for four machine learning algorithms(i.e.,XGBoost,random forest,deep neural network,and K-nearest neighbor)under different feature combinations.Simultaneously,the difference in accuracy between Sentinel-2 SR and Landsat-9 OLI remote sensing images in soil salinization inversion is analyzed.The results show that:(1)The model constructed based on XGBoost algorithm can achieve high-precision prediction of cotton field salinization,with R^(2) higher than 0.74,MSE lower than 0.04,and MAPE lower than 0.13.(2)Under the condition of feature combination 1,Sentinel-2 SR(S3+GBNDVI)and Landsat-9 OLI(SI+NDVI)remote sensing images achieved the highest prediction accuracy using XGBoost algorithm.(3)Sentinel-2 SR image data in cotton field salinization prediction(R^(2)=0.73-0.88)is better than that of Landsat-9 OLI image data.This study realizes the precise monitoring of soil salinization in cotton fields in the Aral Reclamation Area of Xinjiang,which should provide an effective and timely technical reference for soil salinization control and prevention in cotton fields in reclamation areas.

关 键 词：多光谱遥感影像 光谱指数 土壤盐渍化 机器学习模型 阿拉尔垦区 

分 类 号：S562[农业科学—作物学] S156.4