Research On Detection Method Of Peanut Leaf Spot Disease Levels Based On Multi-scale Hyperspectral Reflectance

Peanuts are an important economic crop and a major protein and vegetable oil source.Among the many diseases affecting peanut growth,peanut leaf spot disease is one of the most devastating and can significantly reduce crop yield.Effective and timely detection of peanut leaf spot disease is the key to accurate management in the field.Hyperspectral-based disease detection technology is efficient,objective,nondestructive,and repeatable and is suitable for large-scale disease monitoring,prevention,and control.However,there are differences between disease features at different scales,making the research focus on hyperspectral detection techniques at different scales different.Therefore,this paper detected the severity of peanut leaf spot disease at the leaf,plant,and field scales.Meanwhile,this paper realizes the multi-scale fusion of leaf spot disease detection to explore the transferability and accuracy of spectral features between different scales.This paper’s main research contents and conclusions are as follows.(1)At the leaf scale,this paper utilizes the improved PCA-Loading method,which is based on assigning weights according to contribution rates,to select feature wavelengths for leaf spectra at different disease levels.It constructs three new spectral indexes(SIs)by combining single wavelengths,ratio wavelengths,normalized wavelengths,and their linear combinations.The severity of peanut leaf spot disease is detected using classification methods such as k-nearest neighbor(KNN),support vector machine(SVM),and back propagation(BP)neural network.The experimental results demonstrate that the improved PCA-Loading method achieves higher classification accuracy while utilizing a smaller number of feature wavelengths compared to commonly used methods.Furthermore,the LS-NSI(leaf spots new spectral index)developed in this study exhibits superior performance in detecting the severity of peanut leaf spot disease compared to commonly used SIs,achieving high accuracy in both KNN and SVM,with an overall accuracy(OA)and kappa coefficient(Kappa)of 96.57% and 95.39% respectively.(2)At the plant scale,this paper uses the proposed multi-window-size one-dimensional convolutional neural network(MWS-1D-CNN)+ Attention model to achieve the severity detection of peanut leaf spot disease.The parallel multi-channel 1D-CNN is used to extract spectral features at different scales,and the Attention mechanism is introduced to further focus on the extracted features to achieve accurate detection of peanut leaf spots at the plant scale.The results showed that the proposed MWS-1D-CNN+Attention model obtained high detection accuracy among all tested models,with OA and Kappa of 93.58% and 91.44%,respectively.(3)At the field scale,this study employs the SLIC+HI+D-FCLS method,which combines simple linear iterative clustering(SLIC),homogeneity index(HI),and fully constrained linear least squares based on distance strategy(D-FCLS),to extract crop regions from hyperspectral images.The average spectrum of all pixels within the crop region is considered as the spectral feature for that specific region.The spectral features of all labeled samples are used as inputs to establish a relationship model between the classification methods such as KNN,BP,SVM,and disease levels for the classification and detection of leaf spot disease at the field scale.The results demonstrate that the proposed SLIC+HI+D-FCLS method achieves higher accuracy and operational efficiency compared to other superpixel-based unmixing methods.Additionally,the proposed method outperforms traditional manual approaches.By utilizing the proposed method,a higher detection accuracy can be achieved in KNN,with an overall accuracy(OA)and Kappa coefficient of 89.71% and 86.27%,respectively.(4)This study conducts data analysis,feature wavelength selection,and spectral index construction on hyperspectral reflectance of peanut leaf spot disease at multiple scales to determine the severity of the disease.The results demonstrate that the proposed leaf spot multi-scale spectral index(LS-MSSI)achieves high detection accuracy at different scales,outperforming commonly used spectral indexes.At the leaf and plant scales,MLR(Multiple Linear Regression)exhibits high precision,with an overall accuracy(OA)of 93.77% and 92.50% respectively,and Kappa coefficients of 91.59% and 89.97% respectively.At the field scale,KNN(k-nearest neighbor)demonstrates high accuracy,with an OA of 90.29% and a Kappa coefficient of87.04%.