| Rice(Oryza sativa L.)is the crop served as mian food in the most region of China.The quality and yield are important fators which would influence the security of food market.However,rice usually suffer from some dieases during the sowing to harvesting period due to some controllable or uncontrollable factors,such as variety,temperature,humidity,and fertilizer,etc.As a result,the quality and yield would be seriously affetcted.There are various types of rice diseases,and the symptom of some diseases have similarities,which makes it difficult to be identified correctly.Thus,hyperspectral imaging technology could be applied to major diseases detection of plant.In this paper,rice leaf infected with sheath blight,rice blast and bacterial blight during the seedling stage is taken as the research samples.The main findings are summarized as follows:(1)Traditional machine learning methods and deep learning methods combined with optimal wavelengths selection methods were used to built models for multiple rice diseases(sheath blight,rice blast and bacterial blight)of four rice varieties discrimination.Three traditional machine learning methods,including support vector machine(SVM),and three deep learning methods were established based on full spectra and optimal wavelengths(selected by PCA-loadings,etc.).Among all the models,three CNN models based on optimal wavelengths selected by PCAloadings achieved best results,with the accuracy of both calibration set and prediction set close to100%.It verified CNN models combined with specific optimal wavelengths selection method could be used to identify rice leaves of different varieties infected with mutiple diseases.(2)A rice disease detection model previously applied to the detection of apparently infected samples was applied to the early detection of diseased rice through model transfer.Model was applied to the early detection of infected samples by transfer.The results showed that three CNN models based on optimal wavelengths selected using PCA-loadings in summary(1)had the best performance,with healthy status of all the infected plants identified correctly when there was only light infection syptom at the very beginning.(3)Regression models based on the hyperspectral imaging of eight chemical indicators of rice leaves of different varieties and diseases were established.Combined with the significance analysis,the chemical indicators suitable for the identification of three major rice diseases were selected.The feasibility of detecting major rice diseases using hyperspectral imaging technology has been proved from the microscopic perspective of internal chemical substance changes.Eight chemical indicators targeted in this study,including Chlorophyll a(Chla),etc.These indicators are related to the activity of cells and health status of rice leaves.In this study,tree traditional machine learning methods were used,including partial least squares(PLS).What’s more,two deep learning models were also used for regression.Three optimal wavelengths selection methods were adopted for optimal wavelengths selection,such as successive projections algorithm(SPA).The results show that CAT,MDA and APX are suitable for the detection of sheath blight and bacterial blight.MDA and APX are suitable for the detection of rice blast and healthy samples.The results identified it is feasible to predict chemical indicators inside rice leaves using hyperspectral imaging technology and the selected chemical indicators also could help to discriminate rice leaves with different healthy status. |
