基于Landsat8 OLI影像干旱区绿洲土壤含盐量反演

Inversion of soil salinity of an oasis in an arid area based on Landsat8 OLI images

利用遥感技术进行土壤含盐量的快速检测可为土壤盐渍化治理和绿洲农业合理开发提供科学指导。基于渭干河—库车河三角洲绿洲采集的95个土壤样品,采用光谱指数、波段反射率与实测土壤含盐量,运用多元线性回归、偏最小二乘回归、支持向量机回归和随机森林回归方法构建土壤含盐量估测模型,并利用最优估测结果对研究区土壤含盐量的空间分布格局进行遥感反演。结果表明:通过全子集回归法筛选出与土壤含盐量相关显著的9个光谱因子,相关系数均在0.5以上(P<0.01)。其中盐分指数中SI-T与土壤含盐量的相关系数最大为0.648;对比4种反演模型的估测精度,拟合的效果由高到低依次为随机森林回归>支持向量机回归>偏最小二乘回归>多元线性回归。其中随机森林模型拟合精度表现最佳,训练集和验证集的决定系数分别为0.870和0.766;相对分析误差分别为2.792和2.105,值均大于2,表明模型反演效果较好,有稳定的估测能力;由随机森林模型的反演结果来看,第Ⅰ等级和第Ⅱ等级占比达到41.62%,分布于绿洲内部的耕作区;第Ⅲ,Ⅳ和第Ⅴ等级区共占比56.41%,主要分布在绿洲外围与沙漠的交错带和荒漠区。采用随机森林机器学习建模方法对土壤含盐量进行反演,估测效果明显优于传统的统计模型,可为干旱区绿洲土壤盐渍化监测提供参考。

The rapid detection of soil salinity using remote sensing technology can scientifically guide the soil salinization control and the rational development of oasis agriculture.Based on 95 soil samples from the oasis of the Weigan-Kuqa River delta,this study established four soil salinity estimation models of multiple linear regression,partial least squares regression(PLSR),support vector machine regression(SVR),and random forest regression using the spectral index,band reflectance,and the measured soil salinity.Then,it conducted the remote sensing inversion for the spatial distribution pattern of the soil salinity in the study area using the optimal estimation results.The results are as follows:①Nine spectral factors that were significantly related to soil salinity were screened using the all-subsets regression method,with correlation coefficients of all above 0.5(P<0.01).Among them,the correlation coefficient between salinity index SI-T and the soil salinity was the highest(0.648);②The comparison of estimation precision show that the fitting effect of the four inversion models was in the order of random forest regression>SVR>PLSR>multiple linear regression.Among these models,the random forest model had the best fitting precision.Its training and validation sets had coefficients of determination(R 2)of 0.870 and 0.766,respectively,with relative percent deviation(RPD)of 2.792 and 2.105,respectively,both of which were greater than 2.These results indicate that the random forest model had a good inversion effect and stable estimation capacity;③According to the inversion results of the random forest model,grade I and II zones account for 41.62%and are distributed in the cultivated land area inside the oasis;grade III,IV,and V zones account for 56.41%and are primarily distributed in the desert and the desert-oasis ecotones.Therefore,compared with conventional statistical models,the random forest modeling method can yield significantly better estimation effects in the inversion of soil salinity.This study can be used as a reference for the monitoring of soil salinization in oases in arid areas.