Farmland Surface Soil Moisture Monitoring Method Based On Microwave Optical Multi-source Remote Sensing

Soil is one of the most important natural resources for human survival,which is closely related to many aspects of people’s life.Traditional soil moisture monitoring is mainly based on point measurement,which consumes a lot of labor and labor time,has low work efficiency,and mostly destroys the original structure of the soil to varying degrees,which is not conducive to the measurement in the soil with more vegetation coverage such as farmland.The use of satellite and airborne microwave remote sensing data has the disadvantages of low spatial and temporal resolution,low monitoring accuracy,weather Light and other natural factors have a great impact on the surface vegetation coverage.With the development of precision agriculture and precision irrigation technology,the use of near earth low altitude remote sensing method can quickly,efficiently and accurately realize the task of farmland soil moisture monitoring without contacting and damaging the original soil structure.In this paper,the low altitude scanning method combined with UWB radar is used to collect the microwave remote sensing data of farmland surface soil.The UAV equipped with multispectral camera is used to obtain the multispectral remote sensing image of the target farmland,and the multi-source remote sensing data are fused to realize the task of monitoring the surface soil moisture of vegetation covered farmland.The main research contents and conclusions are as follows:(1)Integrating multi-source remote sensing data and using Support Vector Machines(SVM)model,the classification prediction of surface soil moisture is realized.The research shows that when the model input only uses radar echo data,the overall accuracy of the SVM classification model is 95.59%and the kappa coefficient of the model is 0.9492.When the multispectral data is fused as the model feature input,the prediction effect of the trained SVM model is improved,in which the overall classification accuracy is improved to 98.09%and the kappa coefficient is improved to 0.9780.(2)A Convolutional Neural Networks Regression(CNNR)model is proposed to retrieve farmland surface soil moisture.The traditional machine learning regression model Generalized Regression Neural Network(GRNN)and Support Vector Regression(SVR)are used for comparative analysis.The results show that when five time-domain features and three vegetation index features are selected as the feature input of the regression prediction model,the constructed cnnr model R~2 is 0.9168,RMSE is 0.0089cm~3/cm~3 and RPD is 3.0201,which is the best regression prediction model of soil moisture.After the introduction of multispectral vegetation index information into the three models,the prediction accuracy of the models is significantly improved.It is proved that the fusion of multispectral remote sensing data can reduce the interference of field vegetation cover on radar echo signal and improve the accuracy of soil moisture prediction model.(3)Gaussian white noise is added to the test set signal to analyze the noise robustness of the model.The prediction accuracy of different models under different DB noise is discussed,and the prediction ability of each soil moisture monitoring model is compared.The experimental results show that when the signal-to-noise ratio is 0-30db,cnnr model always has the highest R~2,higher RPD and lower RMSE,and has the best noise robustness among the three models.