A Multi-scale Rice Leaf Blast Detection Method Based On Deep Convolutional Neural Network

Rice leaf blast is one of the most destructive and contagious rice diseases,which greatly impacts the yield and quality of rice and the national economy in China.Therefore,strengthening the rapid detection of rice leaf blasts is significant in guiding disease prevention,control,and food security.The rapid development of spectroscopy and UAV technology provides a favorable opportunity for rapid,non-destructive,efficient,and large-area rice leaf blast detection.However,the existing research on crop disease detection based on spectral technology and remote sensing technology still has problems,such as the lack of in-depth understanding of the spectral response characteristics and change patterns of diseases at different scales,the imprecise mining of the characterization information of diseases at different scales,and the imperfect research on the disease detection models at different scales.Therefore,in this study,the spectral response characteristics and change patterns of leaf,plant,and canopy scale rice leaf blasts and different disease levels were compared and analyzed using rice leaf blasts as the research object.Meanwhile,we researched rice blast detection methods at leaf,plant,and canopy scales using indoor hyperspectral imaging data,outdoor non-imaging hyperspectral data,and UAV hyperspectral remote sensing data,respectively.A method for detecting the degree of rice leaf blast disease based on fused features and deep convolutional neural networks at the leaf scale is proposed.A method for wavelength extraction of spectral features of rice leaf blasts based on deep convolutional neural networks and visualization techniques at the plant scale is proposed.A two-branch deep convolutional neural network-based method for rice leaf blast detection is proposed at the canopy scale.To improve the existing detection capability of rice leaf blasts from different scales and provide some technical support for the accurate detection and timely control of crop diseases.The main research results are as follows.(1)At the leaf scale,the spectral response of rice leaf blast at different disease levels at the leaf scale was compared and analyzed.It was found that the hyperspectral reflectance showed increasing and decreasing response patterns with increasing disease levels in the visible and near-infrared regions,respectively,and a"blue shift"phenomenon appeared in the red-edge region.Meanwhile,the successive projection algorithm(SPA),the random frog method(R-frog),the contour of decision coefficient R~2 and the gray-level co-occurrence matrix(GLCM)were used to screen the wavelengths of spectral features,vegetation indices,and texture features with high correlation with the disease,respectively.And seven deep convolutional neural networks based on single and fused features were used to construct leaf blast detection models,respectively.The results show that the modeling accuracy based on fused features is significantly higher than that of single features.Among them,the deep convolutional neural network that fuses the wavelengths of the spectral features screened by SPA with the texture features obtains optimal detection accuracy.And compared with other detection models such as support vector machines(SVM),the proposed deep convolutional neural network model provides better detection accuracy in disease extent detection with OA and Kappa coefficients of 98.58%and 98.22%,respectively.And in the comparative analysis of model performance,the model in this study has more significant advantages,with a 986 data detection time of 0.22s.Therefore,the spectral feature wavelengths and texture features based on fusion SPA screening can more comprehensively characterize leaf blast disease features.Meanwhile,the deep convolutional neural network-based leaf blast of degree detection model can effectively improve the leaf blast detection accuracy at the leaf scale.(2)At the plant scale,the spectral response characteristics and change patterns of different disease levels of rice leaf blast at the plant scale were analyzed,and it was concluded that the trends of spectral reflectance changes at the plant scale and the leaf scale were in significant agreement.In contrast,the spectral reflectance of different disease levels at the plant scale showed significant variability in the near-infrared region.Also,a deep convolutional neural network(DCNN)structure(Res Net-CBAM)was designed by combining the CBAM attention mechanism with the Res Net model for adequate learning of different classes of disease features.The Guided-Grad CAM visualization method was also used to obtain the spectral feature wavelengths with the highest contribution in the most acquired DCNN.Statistical analysis methods(JM distance,within-class dispersion)were used to perform preliminary validation of the selected wavelengths.They were compared and analyzed with dimensionality reduction methods such as SPA,Competitive Adaptive Re-weighted Sampling(CARS),and R-frog to build random forest(RF)and support vector machine(SVM)detection models,respectively.The results show that the feature wavelengths screened by the Guided-Grad CAM method have good within-class separability and intra-class aggregation,with a JM distance greater than 1.9and within-class dispersion less than 0.4.Meanwhile,compared with the modeling results of the other three downscaling methods,the RF and SVM models constructed by the spectral feature wavelengths screened by the Guided-Grad CAM method achieved the best disease level detection accuracy with OA and Kappa coefficients of 97.21%and 96.55%,96.51%and95.69%,respectively.Therefore,the method proposed in this study can more accurately screen the characteristic spectral wavelengths of different disease levels of rice leaf blasts at the plant scale,effectively improving the detection accuracy.(3)At the canopy scale,the spectral reflectance of rice health,disease,and soil and water bodies and the changing pattern of image characteristics were compared and analyzed.It was concluded that the spectral reflectance of rice plants differed greatly from that of soil and water bodies.The difference in spectral reflectance between healthy and diseased was consistent with the leaf and plant scales.The color characteristics of healthy rice,diseased rice,soil,and water bodies also differed somewhat in the image features,especially at the 99%confidence interval.Meanwhile,a two-branch deep convolutional neural network model was proposed in this study for canopy-scale rice leaf blast detection.The model extracts spatial and spectral features of rice diseases from 2 dimensions,spatial dimension,and spectral dimension,respectively.Meanwhile,the spatial and spectral features are fused to form a combined spatial-spectral feature.A 3D convolutional layer is introduced to learn further the combined spatial-spectral feature,which is used to fully learn the rice disease features and improve the disease detection accuracy.The results show that the model proposed in this study has better detection accuracy with OA and Kappa coefficients of 97.08%and 96.11%,compared with the existing one-dimensional convolutional neural networks(1DCNN),two-dimensional convolutional neural networks(2DCNN)and three-dimensional convolutional neural networks(3DCNN).Also,it has some advantages in the application performance of the model,with a model size of 1.90 M and a detection time of 0.67 s for 1200 data.Therefore,the canopy-scale leaf blast detection model with a two-branch deep convolutional neural network detection model proposed in this study has superior detection capability and application potential.