Study On Early Detection Method Of Greenhouse Cucumber Downy Mildew Based On Hyperspectral Imaging

As a representative crop for greenhouse vegetable cultivation in my country,cucumber has a long growth cycle and many diseases.Cucumber downy mildew is the most common and most serious disease in cucumber cultivation.If it is not detected as early as possible and controlled in time,it will cause devastating losses.Therefore,it is of great significance to carry out research on early detection methods of cucumber downy mildew.Based on hyperspectral imaging,image processing and machine learning technology,this paper studies the early detection method of greenhouse cucumber downy mildew and the extraction method of downy mildew disease characteristic band.The main research contents are as follows:（1）In order to accurately extract the effective disease characteristic band and realize the early detection of cucumber downy mildew,a Dis-CARS-SPA characteristic band combining the improved competitive adaptive reweighting algorithm（CARS）and continuous projection algorithm（SPA）fused with disease difference information is proposed Extraction method and establish an early detection model of cucumber downy mildew.Continuously collect images of cucumber leaves from healthy to seriously diseased day by day,extract the leaf area,calculate the average spectrum of the whole leaf as a sample,and divide it into 7 infection stages according to the number of days and the degree of infection;the downy mildew disease is determined by the envelope elimination method Difference bands,based on disease difference bands,use the optimized CARS algorithm with variable stability loop iteration to extract the characteristic bands from the spectral data of 7 different stages,and then use SPA for secondary dimensionality reduction optimization,and 47 characteristic bands of each stage are combined.With characteristic band data,a least squares-support vector machine（LSSVM）model is established for disease detection.The Dis-CARS-SPA-LSSVM model can achieve a detection and recognition rate of 100% for leaf samples from 2 days of infection to severe disease,and the detection and recognition rate of the test set for 1 day of infection reaches 95.83%,which is compared with the CARS-SPA feature extraction method that does not incorporate disease difference information,it is 4.16 percentage points higher,and the recall rate of infected samples reaches 100%.（2）Extract the samples of the leaf diseased area,perform more accurate feature band extraction,and establish an early disease detection model.The RGB color image of the leaf is balanced by light compensation based on the two-dimensional gamma function illumination unevenness adaptive correction algorithm,and the HSV color threshold segmentation is used to obtain the leaf diseased area at the later stage of the disease as the prior information of the diseased area,and the FAST feature point is used to detect The algorithm extracts the main leaf vein as the reference axis,and performs corner matching based on the prior information of the diseased area to obtain the leaf diseased area at the early stage of the disease.Through the Dis-CARS-SPA feature extraction method,27 feature bands were extracted,and the model was established to obtain 100% detection and recognition rate for samples from 6 stages of infection from 2 days to severe disease,and the recognition rate of the test set for 1 day of infection reached 97.22%,which was 1.39%higher than the model established by the average spectral data of the whole leaf,reduced 20 characteristic bands for modeling,and reduced the model complexity and training cycle.（3）Based on the extracted characteristic bands,the key characteristic wavelengths that can characterize the disease are screened.Through the fusion and comparison of the four proposed methods,the key characteristic wavelengths were screened separately,and the possible band ranges of the key characteristic wavelengths were summarized.Establish a model for the sample of the diseased area,and further simplify the screening to obtain725.4nm and 778.4nm as the key characteristic wavelengths for disease characterization.The recognition rate of the diseased leaves in the diseased area reaches 100%,the recognition rate of the test set for the average spectrum of the whole leaf was 94.68%,and the recognition rate of the test set for the whole leaf for 1 day was 87.5%.