Estimation Of Soil Salt Content In Farmland Based On UAV Remote Sensing

Soil salinization has become one of the most serious land degradation and environmental damage deterioration problems in the world,which is an important factor limiting the sustainable development of agriculture.In order to obtain soil salt content information efficiently and conveniently and realize accurate monitoring of salinization,this paper takes ground measured soil salt data in 2021 and 2022 as the research object,and discusses the feasibility of using UAV multi-spectral remote sensing platform combined with machine learning model to estimate farmland soil salt content in bare soil period and covered soil period.The ground salt data of two periods were collected,and the UAV multi-spectral image data of the study area were obtained synchronously.The spectral bands of the corresponding positions were extracted and the relevant vegetation index and salt index were calculated.The variable screening methods were used to evaluate and analyze the corresponding sensitive spectral variables in the two periods,and combined with different modeling methods,the estimation model of farmland soil salt content in bare soil period and cover period was constructed.Based on the optimal estimation model,the spatial distribution map of soil salt at the best estimated depth(0-20cm)in study area was drawn,and the distribution of soil salinization in the study area was visualized for further analysis and evaluation.The main results of this study are as follows:(1)Study on estimation of farmland soil salt content by UAV remote sensing during bare soil period.The spectral information of surface soil in this period can better reflect its salt content.Pearson correlation coefficient method(PCC)was applied to analyze the correlation between spectral bands and salt index and soil salt content,and some spectral variables(Nir,Red and SI,SI1 and SI3)with strong correlation could be obtained.Then,it is used as the input variable of the model to construct the estimation model.Since the salinity index group is calculated by the combination of different bands,it contains more extensive spectral information.Therefore,compared with the spectral band group,when the salinity index group is used as the input variable,the model has a higher estimation accuracy.Support vector machine(SVM)has the best result among the 3 estimation models,followed by Partial least squares regression(PLSR),and Stepwise regression(SR)has the worst estimation accuracy.SVM model in salinity index group has the best performance,and its validation set R~2,RMSE and MAE are 0.629,0.057 and 0.044,respectively.This indicates that the machine learning regression method introduced is superior to the other two methods,so it is feasible to estimate soil salt content.(2)Estimation of soil salt content in mulch period based on linear variable screening method.Both of the two screening methods can remove relevant jumbled spectral information to improve the efficiency of the model.Compared with Pearson correlation coefficient method(PCC),Grey relation analysis(GRA)is more suitable for screening spectral variables,which plays a great role in improving the estimation accuracy of the model.Through the salt estimation model constructed by two variable screening methods and three machine learning algorithms,it can be seen that Support vector machine(SVM),BP neural network(BPNN)and Random forest(RF)models all achieve good prediction effect to some extent.Considering the stability and estimation accuracy of the models comprehensively,the prediction effect of0-20cm surface soil model is better than that of 20-40cm surface soil model.The R~2,RMSE and MAE of the best model are 0.775,0.055 and 0.038,respectively,which indicates that the selected method has certain applicability in the study of monitoring farmland soil salt content by UAV remote sensing.(3)Estimation and mapping of soil salt content in mulch period based on extreme gradient lifting algorithm.Using Pearson correlation coefficient(PCC)and Extreme Gradient Boosting(XGBoost)method to screen features can remove relevant jumbo-spectral information and improve modeling efficiency.Comparing and analyzing the prediction effect of salt content estimation models constructed based on three modeling schemes,it can be seen that after optimization of variable features of scheme 3:XGBoost-GRA,the overall prediction accuracy and stability of the models have been improved to varying degrees.The prediction effect of 0-20cm soil depth was the best among the three soil depths,and the R~2,RMSE and MAE of the highest accuracy model were 0.820,0.044 and 0.030,respectively.Among the estimation models,the SVM model with the soil depth of 40-60cm had the worst accuracy,and its R~2,RMSE and MAE were 0.437,0.113 and 0.069,respectively.Moreover,the spatial distribution map of 0-20cm soil salinity based on the best estimation model(RF)can reflect the degree of soil salinization in the study area.