Soil Composition Detection Method Based On Frequency Domain Reflection

Soil moisture and salt information are the main components of agricultural information technology,and the important guarantee for achieving precision agriculture.The soil moisture and salt detection sensor based on frequency domain reflection(FDR)technology is widely used in soil moisture and salt information collection due to its advantages of low cost,fast measurement speed,no damage to soil layer and easy operation.Due to the spatial and temporal differences in soil particle size and composition,manual parameter calibration is required frequently in the use of sensors.In order to solve the problems of tedious manual calibration process and long calibration time,this dissertation proposes an automatic parameter calibration method for soil water and salt content detection based on machine learning.The key of sensor parameter calibration is to explore the correlation among the output frequency,soil layer information and soil moisture and salt content.Firstly,based on the dielectric constant characteristics of soil,FDR measurement principle and FDR sensor measurement data of soil moisture and salt content in different time and space,a hybrid model based on machine learning algorithm and FDR measurement mechanism is established.According to the characteristics of the data,the LightGBM algorithm is selected as the modeling algorithm.In the model training,the Bayesian optimization method is used to optimize the LightGBM hyperparameters.The output of LightGBM obtained from training are put into the mechanism model of soil moisture measurement,and the hybrid model of parameter calibration is obtained.The experiment shows that the hybrid model used in this dissertation meets the measurement error of ±5%specified in the "specification for agrometeorological observation" at each measurement point,and its average absolute error is 1.28%.In order to solve the problem of inaccurate prediction of hybrid model caused by spatiotemporal change and soil composition change of the above model in online application,this dissertation introduces the method of transfer learning.According to each region of the probability density distribution of a transfer learning algorithm(TrAdaBoost algorithm)based on instance,improved transfer base learning algorithm with weight updating method,FDR sensor under other spatio-temporal data as secondary data,the current online applications as the source of data space and time domain data,using the improved TrAdaBoost algorithm,only a small amount in the current time and space under the soil sample data to get the accurate FDR sensor calibration model.The experiment shows that the sensor calibration model based on migration learning is more accurate than manual calibration,and it meets the measurement error of ±5%stipulated in the specification of agrometeorological observation,which proves the effectiveness and accuracy of the method.For the establishment of the automatic calibration model of soil salinity,the automatic calibration model of soil moisture is used for reference.The differences between the salinity measurement and moisture measurement mechanism models and the differences in data characteristics are analyzed.The input and output of the model are changed,and the model is re-established.The output conductivity of the model is calibrated to 25℃ based on the relationship between conductivity and temperature.Experiments show that the model error meets the requirements of agricultural norms.Based on different time space FDR sensor data of soil for FDR automatic parameters of soil water and salt measuring sensor calibration.Reducing the conventional manual calibration need to measure the actual soil water-salt information amount of data.Using the measured data in different parts of the experiment,prove that the method can realize precise FDR sensor soil water and salt measuring parameters calibration.