基于无人机多光谱遥感的土壤含盐量反演模型研究

Soil Salt Inversion Model Based on UAV Multispectral Remote Sensing

为探究无人机多光谱遥感技术快速监测植被覆盖下的土壤含盐量问题,以内蒙古河套灌区沙壕渠灌域内4块不同盐分梯度的耕地为研究区域,利用无人机搭载多光谱传感器获取2018年8月遥感影像数据,并对0~40 cm的土壤进行盐分测定。分别引入敏感波段组、光谱指数组、全变量组作为模型输入变量,采用支持向量机(Support vector machine,SVM)、BP神经网络(Back propagation neural network,BPNN)、随机森林(Random forest,RF)、多元线性回归(Multiple linear regression,MLR)4种回归方法,建立基于3组输入变量下的土壤盐分反演模型,并进行精度评价,比较不同输入变量、不同回归方法对模型精度的影响,评价并优选出最佳盐分反演模型。结果表明,通过分析3个变量组的R 2和RMSE,光谱指数组在4种回归方法中均取得了最佳的反演效果,敏感波段组和全变量组在不同的回归方法中反演效果不同。4种回归方法中,3种机器学习算法反演精度明显高于MLR模型,且MLR模型中的敏感波段组和全变量组均出现了“过拟合”现象,RF算法在3种机器学习算法中表现最优,SVM算法和BPNN算法在基于不同变量组的模型中表现也不相同。基于光谱指数组的RF的盐分反演模型在12个模型中取得了最佳的反演效果,R^2 c和R^2 v分别达到了0.72和0.67,RMSEv仅为0.112%。

Fast acquisition of soil salt content under vegetation cover is the objective need of saline soil management and utilization.Four kinds of croplands with different salinization values in Shahaoqu District of Hetao Irrigation Area were set as the study areas.The UAV equipped with a multi-spectral camera obtained the remote sensing image data of August,meanwhile,the soil salinity with depth of 0~40 cm was tested.The sensitive band group,spectral index group and full variable group were introduced as model input variables.Four regression methods,including support vector machine(SVM),BP neural network(BPNN),random forest(RF)and multiple linear regression(MLR),were used to establish soil salinity inversion models which were based on three groups of input variables,respectively.Firstly,the model precision was evaluated,and then the effects of different input variables and different regression methods on the model precision were compared,finally the best salt inversion model was evaluated and optimized.The results indicated that comparing the R 2 and RMSE of three variable groups,the spectral index group achieved the best inversion effect between the four regression model methods,and the sensitive band group and the full variable group had advantages and disadvantages in different regression algorithms.Between the four regression methods,the inversion accuracy of three machine learning regression algorithms was significantly higher than that of the MLR model.Moreover,both the sensitive band group and the full variable group in the MLR model showed the phenomenon of“overfitting”.And RF algorithm performed best between the three machine learning algorithms.Besides,SVM algorithm and BPNN algorithm performed better and worse in the model with different variable groups.The RF salt inversion model based on the spectral index group achieved the best inversion effect among the 12 models,the R^2 c and R^2 v reached 0.72 and 0.67,respectively,and the RMSE v error was only 0.112%.The research result can provide a theoretical reference for soil salinity monitoring in arid and semi-arid areas.