Fast Detection Methods For Disease Early Diagnosis And Physiological Information Determination Of Tomato

Precision agriculture and digital agriculture are the most frontier technologies in modern agriculture, and they are also the key and kernel technologies for the development of modern agriculture and the realization of sustainable agriculture. Precision agriculture and digital agriculture require the fast, accurate, digital and positional agricultural production and management information. However, the traditional lab and chemical measurements cannot fulfill the fast, accurate, dynamic and high efficient demand of modern agriculture. Therefore, this study is mainly focused on the tomato (Lycopersicum esculentum), and aims to develop the fast and accurate disease early diagnosis methods and physiological information detection methods of tomato under different disease stress. This study will supply a new approach for the precision management and disease prevention and treatment of tomato, which is also meaningful for the precision production and plant of tomato. The main results were achieved as follows:(1) The spectral recognition models and imaging recognition models were developed for the early diagnosis of gray mold disease of tomato stems. An analysis routine was developed for data preprocessing, effective and feature information extraction, linear and nonlinear recognition models, which was effective for the disease early diagnosis of tomato stems. Different spectral preprocessing methods were compared for full-spectrum partial least squares (PLS) models. Effective wavelengths (EWs) were selected by loading weights and used as input of recognition models. The optimal recognition ratio was achieved by effective wavelength-least squares-support vector machine (EW-LS-SVM) model with a correct recognition ratio of100%for validation set. Probability statistics filter and2-deravertive probability statistics filter were used for texture feature extraction of hyper-spectral imaging information. The PLS model obtained a correct recognition ratio of97.37%for validation set. Genetic algorithm-partial least squares (GA-PLS) method was applied to extract the effective texture features, and PLS and SVM models were developed for gray mold disease diagnosis of tomato stems. The optimal model was GA-PLS-PLS model and the correct recognition ratio was92.11%for validation set.(2) A synchronous diagnosis model using spectral information was developed for three disease of tomato, including gray mold disease, sclerotinia sclerotiorum disease and early blight disease. The visible and near infrared (400-900nm) spectral information was extracted for hyper-spectral imaging data, and different spectral preprocessing methods were compared for the development of PLS and extreme learning machine (ELM) models. The optimal model was full-spectrum ELM model (Detrending) with a correct recognition ratio of94.20%for synchronous diagnosis of three diseases. GA-PLS was applied for EW selection, and then PLS, back propagation neural networks (BPNN), SVM and ELM models were developed for disease diagnosis. The optimal results were achieved by ELM model with a correct recognition ratio near90%for validation set.(3) A synchronous diagnosis methods for three diseases and each of two kinds of diseases were developed using hyper-spectral imaging information. GA-PLS was used to settle the effective wavelength, and the corresponding imaging of these effective wavelengths were selected for texture feature extraction using probability statistics filter and2-deravertive probability statistics filter. Four kinds of models were developed for disease recognition, including PLS, BPNN, SVM and ELM. The results indicated that the correct recognition ratio for the combination of gray mold, early blight diseases and healthy leaves, and the combination of gray mold, sclerotinia sclerotiorum diseases and healthy leaves were over90%, which were acceptable results. However, the recognition ratio for the combination of sclerotinia sclerotiorum, early blight diseases and healthy leaves, the combination of these three diseases and healthy leaves were both less than80%.(4) The physiological information of peroxidase (POD) of tomato leaves under gray mold disease stress was determined using spectral technology. A comparison of full-spectrum PLS and ELM models were proceeded, and the results indicated that full-spectrum ELM model (MSC) achieved a better prediction performance with r=0.8297and root mean squares error of prediction (RMSEP)=983.7830for validation set. Moreover, GA-PLS was used to extract effective wavelengths, and21EWs were selected to develop GA-PLS-PLS, GA-PLS-MLR, GA-PLS-ELM models to determine POD of tomato leaves. The results indicated that GA-PLS-ELM model (SG) achieved an optimal prediction performance with r=0.8647and RMSEP=465.9880for validation set, which is also the best result for all developed models. The overall results demonstrated that it was feasible to detect the POD of tomato leaves under gray mold disease stress using spectral technology, which supplied a new approach for physiological parameter detection of tomato leaves.