Non-destructive Detection Of Rapeseed Growth And Quality Information Based On Spectral And Data Mining Technology

Precision agriculture is the developing trend of global agriculture in the 21st century. The technologies of crop growth information collection, information management and variable operations are the key technologies of precision agriculture. Fast and precise collection of the crop growth information is one of the important basis, hot and difficulties of precision agriculture.Considering the problems and deficiencies of current research of precision agriculture, this thesis was focus on rapeseed (Brassica napus). Quadratic regression orthogonal design was applied in the field experience to produce different levels of nitrogen (N), phosphorus (P), potassium (K) and boron (B) of rapeseed. Visible and near infrared spectroscopy technology was applied to study the relationship between reflectance spectra and the content of N, P, K, and B of rapeseed leaf and canopy. The relationship between N content of rapeseed leaf and SPAD value, and relationship between 3CCD multi-spectral image of rapeseed canopy and SPAD value were also investigated in this thesis. Furthermore, the influence of gamma-ray treatment on spectral characteristic of rapeseed was studied. The main research achievements were as follows:(1) The relations between reflectance spectra and content of N, P, K, and B content of rapeseed leaf were studied. Chemometric models were built by the full waveband and optimal wavelengths. The results showed that the best models of the four elements were all developed by direct orthogonal signal correction (DOSC) method combined with partial least squares (PLS) method. The prediction coefficients of the four best models were 0.9743,0.6971,0.9316 and 0.8903 for N, P, K and B, respectively. Successive projections algorithm (SPA) method was applied to select the optimal wavelengths with least collinearity and redundancies. The optimum wavelengths selected by DOSC combined with SPA method (DOSC-SPA) were 958,540,627 and 686 nm. The best model of N content was built by DOSC-SPA combined with least squares-support vector machine (LS-SVM) method. The correlation coefficient for prediction set was 0.9737. The best models of P, K, and B content were all built by DOSC-SPA combined with back-propagation neural network (BPNN) method. The correlation coefficients for prediction set were 0.7054,0.9380 and 0.8916 for P, K and B, respectively.(2) The relations between reflectance spectra and content of N, P, and K content of rapeseed canopy were studied. Chemometric models were built by the full waveband and optimal wavelengths. After comparing prediction results of the PLS models built by different pre-processing methods, DOSC was chosen as the best pre-processing method of these three elements. The correlation coefficients for prediction set were 0.9440,0.8260 and 0.9574 for N, P and K, respectively. The optimal wavelengths selected by DOSC-SPA were 761,994 and 927 nm, respectively. The best models of N and P content were both developed by DOSC-SPA-LS-SVM. The correlation coefficients for prediction set were 0.9423 and 0.8124 for N and P, respectively. The best model of K content was built by DOSC-SPA-PLS, and correlation coefficient was 0.9526.(3) The relation between N content of rapeseed leaf and SPAD value was studied. The result indicated that there existed a linear relationship between the N content and SPAD value during the growth period. The coefficient was 0.861. The prediction precisions of the three unknown samples were 78.8%,91.8% and 94.12%.(4) The relation between multi-spectral image of rapeseed canopy and N content was studied. Multi-spectral image test system which combined visible and near infrared spectroradiometer with 3CCD camera was set up. The models between SPAD value and NDVI, GNDVI and Ratio which were based on multi-spectra reflectance of rapeseed canopy were built. The correlation coefficients for prediction set were 0.932,0.918 and 0.885 for NDVI, GNDVI and Ratio, respectively.(5) The changing rule of the spectral characteristic of rapeseed after being treated by gamma-ray was studied. The influences of different pre-processing combinations and nodes number of hidden layers to the models were discussed. As a result, the optimal model was established and the parameters of the model were shown as follows. The original spectra data were pretreated by smoothing media filter, multiplicative scatter correction and 2nd Savitzky-Golay derivatives. The 6 PLS principal components were transformed by using natural logarithm transformation method. The nodes number of hidden layers of the BPNN model was selected as 4 or 9. The results indicated that after being treated by gamma-ray, the spectral characteristic of rapeseed would change greatly. The prediction precision of the optimal model to distinguish the untreated samples from gamma-ray treated samples was 100%. The precision of predicting the dosages of gamma-ray treatment of all samples achieved 85.71%. It can be concluded that the visible and near infrared spectroscopy could be used to estimate the influence of different gamma-ray dosages on the spectral characteristic of treated rapeseed.