Research On Soil Moisture Content Inversion Model Based On UAV Multi-spectral Remote Sensing

Soil moisture is closely related to global climate change,carbon cycle,water cycle,agricultural production,ecological protection and restoration.In view of this,this paper takes bare land,alfalfa covered land and wheat covered land in the irrigation area of Taolai River in Jiuquan City,Gansu Province as the research area,obtains soil moisture content information and UAV multispectral remote sensing images,studies the correlation between spectral reflectance,spectral index and soil moisture content,and uses support vector machine,BP neural network,random forest and partial least squares four machine learning algorithms to construct soil moisture content inversion model.During the vegetation coverage period,the random forest,BP neural network,support vector machine and partial least squares method were used to construct the soil moisture content estimation model based on variable importance analysis algorithm,gray correlation analysis algorithm and continuous projection algorithm.Finally,three accuracy evaluation indexes were used to comprehensively evaluate the inversion model.The main conclusions are as follows:(1)Based on the spectral reflectance,the inversion models of soil moisture content in bare land,alfalfa covered land and wheat covered land were constructed.The results showed that among the five-band spectral reflectance extracted from UAV multispectral remote sensing images,the red band,green band,blue band,near-infrared band and red band all showed significant correlation with soil moisture content,and the red band had the strongest correlation with soil moisture content.In the model inversion,the R~2 of the modeling set and the verification set of the soil moisture content inversion model based on the band spectral reflectance is mostly distributed between 0.6 and 0.9,the root mean square error RMSE is less than 2,and the relative error RE is below 20%.All models have the ability to quantitatively estimate the soil surface moisture content,indicating that the UAV multi-spectral remote sensing can effectively invert the soil surface moisture content.The accuracy of soil moisture inversion models constructed by four machine learning algorithms was comprehensively compared.For bare land,the best inversion model for soil moisture inversion based on spectral reflectance was RF,and the best inversion model for alfalfa cover and wheat cover was SVM.(2)Based on the spectral index,the inversion model of soil moisture content in alfalfa and wheat covered land was constructed.Different spectral indices are obtained by combining and transforming the spectral reflectance extracted from UAV multi-spectral remote sensing images.The correlation between 18 spectral indexes and soil moisture content corresponding to alfalfa coverage and wheat coverage was analyzed,and the inversion model of soil moisture content based on spectral index was constructed.The results showed that among the four machine learning algorithm models,SVM,BPNN and PLSR had higher accuracy than RF algorithm model,and R~2 reached more than 0.548.RF model had certain limitations in dealing with soil moisture content and vegetation index,and R~2 reached more than 0.537.Comparing and analyzing the eight soil moisture content inversion models,the validation set R~2 of the SVM inversion model based on vegetation index in wheat coverage reached 0.827,RMSE was 0.893,and RE was 9.7%,which was the optimal value in the eight models.Therefore,the SVM model based on vegetation index is the best soil moisture inversion model.(3)Variable screening method combined with machine learning algorithm to invert soil moisture content.The variable importance projection(VIP),gray relational analysis(GRA)and successive projections algorithm(SPA)screening methods were used to determine the modeling factors of the inversion model of soil moisture content in alfalfa-covered land.By comparing and evaluating the inversion effects of each model before and after screening,the results showed that the use of variable screening methods can significantly improve the prediction accuracy of the model.VIP is the best screening method for estimating soil moisture content by UAV multispectral remote sensing under this vegetation cover condition.The validation set R~2 of each machine learning model can reach more than 0.65,and RMSE is about 1.1.Under this screening method,the overall fitting effect of the model is good,and the inversion error is small.Through comprehensive evaluation,the three screening methods of UAV multispectral inversion of farmland soil moisture content under vegetation coverage are VIP,GRA and SPA.In addition to the selection of variable screening methods,machine learning algorithms also have a great impact on the accuracy of soil moisture content estimation models.Among the 12 inversion models,the VIP-SVM model has the highest estimation accuracy(R~2=0.7993)and shows stronger robustness,followed by BPNN model and RF model.