The Rapid Detection Of Soil Moisture Based On Near-infrared Spectroscopy And Machine Vision

Soil moisture has a direct impact on crop growth, agricultural micro-climate, as well as soil mechanical properties. Soil moisture is also an important parameter in agriculture, water and soil conservation, and soil erosion. Accurate testing soil moisture is vital for water resources management, irrigation, crop growth, water-saving, production forecasts, as well as chemicals monitoring, and also is a key technology in precision agriculture. Meanwhile, the detection of soil surface moisture plays an important role in effective distribution of energy, infiltration and run off, as well as soil erosion dynamic processes, such as soil erosion forecasting. Therefore, it is quite important to test soil moisture rapidly and accurately.The methods of soil moisture detection based on near-infrared spectroscopy and machine vision were studied. First, near-infrared spectroscopy was introduced to analyze the soil water content. Three kinds of soil from Hubei province and the loessial soil from the Loess Plateau region were collected for analysis, in which 10 kinds of spectral pretreatment methods for modeling were compared. The Fourier feature extraction of near-infrared spectroscopy was used to optimize soil moisture analytical model; the soil testing system based on machine vision with both software and hardware was constructed; machine vision was used to build the soil moisture analytical model. In order to improve the analytical results, data fusion technology of machine vision and near-infrared spectroscopy was taken to build soil moisture testing method. The main results are as follows:1. Four kinds of soil samples were used, three of which were collected from Hubei province, namely paddy soil, yellow brown soil and alluvial soil, and the fourth is the loess soil, which was collected from the Loess Plateau Region. All samples were only prepared with a minimum coarse sieving and air drying to remove any miscellany.2. The 3 kinds of soils collected from Hubei province were taken for study, qualitative and quantitative models for soil moisture content was set up based on near-infrared spectroscopy. Principal component analysis was used to classify the different moisture content; The decision coefficient R~2 of the near-infrared spectroscopy quantitative model was 0.9946, with a standard error of 0.801% for the interactive validation model, and the model forecast decision coefficient R~2 was 0.9919, with the forecast standard error 0. 912 %.3. when Kennard-Stone algorithm was used to screen out samples to expand the sample space, the untreated soil samples with different water content were predicted by re-establishment of PLS soil moisture prediction model, and the error can reduce from 0.78% to 0.42%, which showed that soil samples space expansion method is capable of accuracy improvement.4. The method of slope and intercept revising was used to correct the raw samples model forecast value, the error can reduce from 0.78% to 0.38%, and the result indicated that the slope/intercept revising can enhance the compatibility of the testing model.5. Wavelet was used to eliminate noise of spectrum before modeling; the result showed that the wavelet could effectively remove the noise information of the spectrum.6. Parameters extraction on Fourier feature of near-infrared spectroscopy was used to analyze the soil moisture. The 3 kinds of soils collected from Hubei province were taken for study, when 15 parameters were choosen, the PLS prediction model, collected from Fourier transform parameters and characteristics of soil moisture, could be optimal, and the model decision coefficient and RMSECV were 0.9881 and 1.106%, respectively; the model forecast decision coefficient R~2 and the forecast standard error were 0.9811 and 1.185 %, respectively. The prediction error of loessial soil from the Loess Plateau region could reduce from 4% to 2% compared with the traditional model, and the accuracy was improved greatly.7. The system of soil surface image acquisition hardware based on machine vision was constructed, and the soil image feature extraction system under Matlab GUI platform was developed.8. The experiment of machine vision technology was introduced to analyze soil moisture. A linear regression model between gray level of soil image and the soil moisture was established, the model's decision coefficient was 0.784. It was confirmed that the value of the gray level and the soil moisture content have good correlation; The multi-linear model and the BP neural network nonlinear model, between characteristics of the image parameters and soil moisture,were established. H, S and V were used to establish the non-linear neural network model, which was optimal, and when the hidden layer was 10, the model decision coefficient could reach 0.9849.9. The technology of data fusion based on near-infrared spectroscopy and machine vision was used to detect soil moisture, in which data were integrated at the features level. The feature of near-infrared spectral and the parameters HSV of the soil image were taken to build data fusion model. Artificial neural network technology was used for data integration, and the BP and RBF model were established ultimately. The decision coefficient of the BP data fusion model was 0.9961, which indicates that the forecast accuracy was significantly improved. Besides, the BP model's accuracy was higher than that of RBF model, but RBF was faster than BP model.