SOIL PROPERTIES PREDICTION FOR REAL-TIME SOIL SENSOR BASED ON NEURAL NETWORK

摘要

Real-time soil sensor (RTSS) was built and tested, making use of a near-infrared spectrophotometer, which offered a convenient and quick method for in-situ soil organic matter (SOM), total nitrogen (TN), pH and moisture content (MC) measurement. The sensor could collect a spectrum absorbance of soil (i.e. 500-1650nm with 7nm interval). Neural network was used for making prediction model for each soil component. The training and testing of neural network was based on 1300 dataset was taken by the RTSS from 7 location of Japan where included paddy and upland crop field. Input variables represent spectrum absorbance of the points of interest, while the output variables represent SOM, TN, pH and MC data of the points of interest, which was analyzed in the laboratory. After the neural network has been successfully trained, its performance was tested on a separate testing set. The result of MC, pH, SOM and TN prediction indicated that the NN model validated coefficient of determination of R{sup}2=0.91, 0.75, 0.95 and 0.96, respectively.