摘要
基于野外采集的110个表层土壤盐分样本,采用四种机器学习算法(随机森林、极限学习机、多元自适应样条回归和人工神经网络)与46个环境协变量来绘制新疆艾比湖地区土壤盐分的空间分布图并预测了其不确定性.通过10折交叉验证对各模型的精度进行比较,研究结果表明:(1)使用随机森林(Random Forest,RF)模型能够更稳定更准确地预测土壤EC值,其R^(2)均值达到了0.662,该性能优于人工神经网络(0.622)、极限学习机(0.637)和多元自适应样条回归模型(0.549).(2)Sentinel-2光谱数据是土壤EC预测最重要的变量,其次是盐分指数、气候、地形数据以及植被指数,相对重要性分别为44%、31%、20%和5%.(3)RF模型的结果揭示了区域土壤盐分空间分布的变化信息,模拟结果精度优于其余三个模型,确定RF模型是干旱区尾闾湖流域土壤盐分监测的有效方法.
Based on 110 topsoil salinity samples collected in the field,this study used four machine learning algorithms(Random Forest,Extreme Learning Machine,Multivariate Adaptive Regression Splines and Artificial Neural Network)and 46 environmental covariates to map the spatial pattern of soil salinity(EC)and predicted its uncertainty in the Ebinur Lake area,Xinjiang,China.The precision of each model was compared through 10-fold cross validation.The results indicated that:(1)The Random Forest model could predict the soil EC value more stably and accurately,and its average R^(2) reached 0.662;The performance was better than that of Artificial Neural Network(0.622),Extreme Learning Machine(0.637)and Multivariate Adaptive Regression Splines model(0.549).(2)The results revealed that Sentinel-2 spectral data was the main variable for soil EC prediction,followed by salinity index,climate,terrain data and vegetation index,with relative importance of 44%,31%,20%and 5%respectively.(3)In addition,the map based on the RF model reveals the most reasonable change information in the spatial distribution of EC,while the other three models have produced some errors on the salinization degree of the region.Therefore,this study determined that the RF model is an efective method for monitoring soil salinity in the terminal lake of the arid region.
作者
王瑾杰
丁建丽
张子鹏
张喆
WANG Jinjie;DING Jianli;ZHANG Zipeng;ZHANG Zhe(School of Geographical and Remote Science,Xinjiang University,Urumqi Xinjiang 830017,China;Xinjiang Key Laboratory of Oasis Ecology,Urumqi Xinjiang 830017,China;Key Laboratory of Smart City and Environment Modeling of Higher Education Institute,Urumqi Xinjiang 830017,China)
出处
《新疆大学学报（自然科学版）（中英文）》
CAS
2022年第5期513-521,529,共10页
Journal of Xinjiang University(Natural Science Edition in Chinese and English)
基金
国家自然科学基金联合重点项目(U2003202)
国家自然科学基金(41961059)
新疆维吾尔自治区自然科学基金重点项目(2021D01D06)
新疆维吾尔自治区科技厅重点实验室开放课题(2020D04038)
新疆维吾尔自治区教育厅高校科研计划(XJEDU2021Y009)。
