Prediction and monitoring model for farmland environmental system using soil sensor and neural network algorithm

článek v časopise ve Web of Science, Jimp

en

In this study, data fusion algorithm is used to classify the soil species and calibrate the soil humidity sensor, and by using edge computing and a wireless sensor network, farmland environment monitoring system with a two-stage calibration function of frequency domain reflectometer (FDR) is established. Edge computing is used in system nodes, including the saturation value of the soil humidity sensor, the calculated soil hardness, the calculation process of the neural network, and the model of soil classification. A bagged tree is adopted to avoid over-fitting to reduce the prediction variance of the decision tree. A decision tree model is established on each training set, and the C4.5 algorithm is adopted to construct each decision tree. After primary calibration, the root mean squared error (RMSE) between the measured and standard values is reduced to less than 0.0849%. The mean squared error (MSE) and mean absolute error (MAE) are reduced to less than 0.7208 and 0.6929%. The bagged tree model and backpropagation neural network are used to classify the soil and train the dynamic soil dataset. The output of the trained neural network is closer to the actual soil humidity than that of the FDR soil humidity sensor. The MAE, the MSE, and the RMSE decrease by 1.37%, 3.79, and 1.86%. With accurate measurements of soil humidity, this research shows an important guiding significance for improving the utilization efficiency of agricultural water, saving agricultural water, and formulating the crop irrigation process.

In this study, data fusion algorithm is used to classify the soil species and calibrate the soil humidity sensor, and by using edge computing and a wireless sensor network, farmland environment monitoring system with a two-stage calibration function of frequency domain reflectometer (FDR) is established. Edge computing is used in system nodes, including the saturation value of the soil humidity sensor, the calculated soil hardness, the calculation process of the neural network, and the model of soil classification. A bagged tree is adopted to avoid over-fitting to reduce the prediction variance of the decision tree. A decision tree model is established on each training set, and the C4.5 algorithm is adopted to construct each decision tree. After primary calibration, the root mean squared error (RMSE) between the measured and standard values is reduced to less than 0.0849%. The mean squared error (MSE) and mean absolute error (MAE) are reduced to less than 0.7208 and 0.6929%. The bagged tree model and backpropagation neural network are used to classify the soil and train the dynamic soil dataset. The output of the trained neural network is closer to the actual soil humidity than that of the FDR soil humidity sensor. The MAE, the MSE, and the RMSE decrease by 1.37%, 3.79, and 1.86%. With accurate measurements of soil humidity, this research shows an important guiding significance for improving the utilization efficiency of agricultural water, saving agricultural water, and formulating the crop irrigation process.

humidity sensor; sensor calibration; backpropagation; bagged tree; neural network; IOT; DROUGHT; DESIGN; CONDUCTIVITY; AGRICULTURE; TEMPERATURE; MANAGEMENT; IMPACTS; CLIMATE

04.02.2023

DE GRUYTER POLAND SP Z O OBOGUMILA ZUGA 32A STR, 01-811 WARSAW, MAZOVIA, POLAND

DE GRUYTER POLAND SP Z O OBOGUMILA ZUGA 32A STR, 01-811 WARSAW, MAZOVIA, POLAND

2391-5471

21

1

17