Early Warning And Identification Of Rice Sheath Blight Based On Multi-source Data Fusion

With the expansion of rice planting areas,early warning and identification of rice diseases have become increasingly important.Traditional methods relying solely on meteorological forecasts are too macroscopic,and detection of pathogens is inefficient and requires professional technical personnel.While time series analysis and spectroscopy provide possibilities for high-throughput and rapid detection of rice diseases.This paper focuses on sheath blight and achieves multi-level warning and improved accuracy of early warning and identification through multi-source data fusion.Firstly,monitoring and early warning are carried out based on meteorological data.After reaching the warning level,physiological and biochemical information of leaves is analyzed through spectral data.In addition,disease area recognition and disease grade classification are conducted using disease image recognition technology to achieve early identification of the disease in rice fields.The main research contents are as follows:(1)Research on early warning of rice sheath blight based on meteorological factors.Meteorological data is used for the first macroscopic warning of rice sheath blight,and after reaching different warning levels,different operations are combined for accurate disease warning.Temperature and relative humidity are selected as the key factors for sheath blight meteorological warning.The CNN-GRU-Attention model has high prediction accuracy and the shortest training time in its 6-hour forecasting task.For temperature prediction,R2is 0.975 with a training time of107.94 seconds,while for humidity prediction,R2is 0.942 with a training time of 109.34 seconds.The meteorological warning level is also divided.(2)Research on early warning of rice sheath blight based on physiological information of leaves.Net photosynthetic rate,stomatal conductance,maximum photochemical efficiency,and effective photochemical quantum yield are selected as physiological warning indicators,and comprehensive modeling analysis is carried out in combination with hyperspectral data.The study found that there were significant differences between the experimental group and the control group during the diseased stage,and the overall trend of each indicator in the diseased experimental group was a sharp decrease after infection,followed by a slow local increase,but with different inflection points.Combined with hyperspectral modeling analysis,the optimal preprocessing methods are determined to be CWT+MMS,CWT+MSC+SG+SS,and CWT+SS.The continuous projection algorithm is selected as the feature extraction method,and the HHO-KELM modeling method is used.The MSE of the optimal prediction models for the four physiological indicators in the prediction set are 2.435,2.530,2.678,and 2.591,and the R~2are0.812,0.743,0.798,and 0.689,respectively.(3)Research on early warning of rice sheath blight based on biochemical information of leaves.Phenylalanine ammonia-lyase,superoxide dismutase,malondialdehyde,andβ-1,3-glucanase are selected as biochemical warning indicators for sheath blight,and comprehensive modeling analysis is carried out in combination with hyperspectral data.The study found that during the diseased stage,there were significant differences in the activity of the four biochemical indicators between the experimental group and the control group.The trend of each indicator in the diseased experimental group was generally a rapid increase followed by a slow decline or fluctuation,and it remained at a relatively high level.The hyperspectral modeling analysis is similar to that of the physiological indicators,and the MSE of the optimal prediction models for the four biochemical indicators in the prediction set are 2.654,2.548,2.746,and 2.604,and the R~2are 0.822,0.854,0.816,and 0.829,respectively.(4)Early identification and grading of sheath blight based on deep learning.After dividing the rice sheath blight image data into disease severity levels and enhancing the data,three convolutional neural networks,VGG-16,Mobile Net V1,and Mobile Net V2,were used for early identification and grading of sheath blight.From the performance of the training set and the test set,the lightweight convolutional neural network Mobile Net V2 model performed the best with an accuracy of 0.9592,and also had the shortest single training time,averaging 23.62 seconds.(5)Structural framework design of rice sheath blight early warning and identification system.Based on the above analysis,the system structure,functional module design,and early multisource data fusion analysis process are comprehensively designed.