达拉特旗黄河南岸盐碱化土壤不同含盐量估算模型对比  

Comparison of Different Salinity Estimation Models for Salinized Soils on South Bank of Yellow River in Dalat Banner

作　　者：刘霞[1] 胡宇 张圣微[1] 白燕英[1] 张欢 LIU Xia;HU Yu;ZHANG Shengwei;BAI Yanying;ZHANG Huan(College of Water Conservation and Civil Engineering,Inner Mongolia Agricultural University,Hohhot 010018,China)

机构地区：[1]内蒙古农业大学水利与土木建筑工程学院,呼和浩特010018

出　　处：《农业机械学报》2024年第10期360-370,共11页Transactions of the Chinese Society for Agricultural Machinery

基　　金：内蒙古自治区科技计划项目(2021GG0369);国家自然科学基金项目(52369009);鄂尔多斯市科技局项目(2021EEDSCXSFQZD01);内蒙古自然科学基金项目(2023MS05024)。

摘　　要：内蒙古自治区鄂尔多斯市达拉特旗黄河南岸由于气候干旱,降水量少,年蒸发量远大于年降水量,靠近黄河地下水位高,导致土壤盐渍化问题突出。以达拉特旗黄河南岸盐碱地为研究对象,基于Sentinel-1、Sentinel-2、Landsat-8和SRTM DEM多源数据,采取相关性分析和连续变量投影结合索套回归(Lasso)、随机森林回归(Random forset,RF)、轻量梯度提升机模型(Light gradient boosting machine,LightGBM)、极端梯度提升模型(Extreme gradient boosting,XGBoost)、一维卷积神经网络(One dimensional convolutional neural networks,1DCNNs)、深度神经网络(Deep neural network,DNN)6种模型进行春季裸土期与植被覆盖期土壤含盐量估算。结果表明:XGBoost模型精度最高,春季裸土期、植被覆盖期测试集决定系数(R2)为0.76、0.58;均方根误差(RMSE)为5.76、7.22 g/kg;平均绝对误差(MAE)为3.38、4.33 g/kg。多源遥感数据结合变量筛选方法利用XGBoost模型揭示研究区不同季节土壤盐分空间分布最有效,含盐量反演结果与野外实际调查分析结果基本吻合。变量重要性分析表明春季裸土期、植被覆盖期重要反演因子分别为:盐分指数(48.3%)、地形因子(33.8%);植被指数(22%)、地形因子(47.9%)。本研究为达拉特旗黄河南岸盐碱地遥感反演提供了有效方法,为春季裸土期与植被覆盖期盐碱化土壤监测及预防提供了理论依据。The south bank of the Yellow River in Dalate Banner,Ordos City,Inner Mongolia Autonomous Region,is characterized by arid climate,low precipitation,annual evaporation much larger than annual precipitation,and the proximity to the Yellow River leads to a high water table,which leads to prominent soil salinization.Taking the saline soil along the south bank of the Yellow River in Dalate Banner as the research object,based on the multi-source data of Sentinel-1,Sentinel-2,Landsat-8 and SRTM DEM,correlation analysis and continuous variable projection combined with Lasso regression(Lasso),random forest regression(RF),light gradient boosting machine model(LightGBM),extreme gradient boosting(XGBoost),one dimensional convolutional neural networks(1DCNNs),and deep neural network(DNN)were used to estimate soil salinity during spring bare soil period and vegetation cover period.The results showed that the XGBoost model had the highest accuracy,and the coefficients of determination(R2)of the test sets were 0.76 and 0.58 for the spring bare soil period and vegetation cover period,the root mean square errors(RMSE)were 5.76 g/kg and 7.22 g/kg,and the mean absolute errors(MAE)were 3.38 g/kg and 4.33 g/kg.The combination of multi-source remote sensing data and the variable screening method by using the XGBoost model revealed that the soil salinity spatial distribution in different seasons in the study area was the most effective,and the results of salinity inversion basically coincided with the results of the actual field investigation and analysis.The variable importance analysis showed that the important inversion factors in the spring bare soil period and vegetation cover period were salinity index(48.3%)and topography factor(33.8%),vegetation index(22%)and topography factor(47.9%),respectively.The research result can provide an effective method for remote sensing inversion of saline and alkaline land on the south bank of the Yellow River in Dalat Banner,and provide a theoretical basis for monitoring and preventing salinized so

关 键 词：干旱盐碱地 土壤含盐量估算 机器学习模型 深度学习模型 

分 类 号：S127[农业科学—农业基础科学]