Research On Wheat Head Blight Detection Based On New Spectral Index

Fusarium head blight(FHB)is a kind of wheat disease that is highly contagious and has a high incidence.It may occur from the seedling stage to the ear stage.Therefore,timely monitoring of disease development during the wheat growth period and accurate identification of diseased kernels after wheat is mature are essential for ensuring the yield and quality of wheat in China.Traditional detection of head blight mainly includes manual visual inspection and chemical and biological experiments,which has problems such as time-consuming,labor-intensive and high cost.We use hyperspectral technology to design two new spectral indices to study FHB on two scales of wheat ears and grains.The main research contents are as follows:(1)For the wheat ears in the growing period,use the SOC710E imaging spectrometer to collect the spectrum data,analyze the data of three different years and varieties in the 400-1000nm spectral range,and propose a universal New Spectral Disease Index(NSDI).First,Gradient Boosting Decision Tree(GBDT)was selected to screen the characteristic bands of samples of each year,and NSDI was formed in the form of normalized wavelength difference.After further characteristic bands optimization,a new NSDI was constructed after confirming the best single wavelength in each year.The predicted coefficient of determination(R~2)of disease severity of each year is above 0.91,and the root mean square error(RMSE)is lower than 0.069.In actual industrial production,fine narrowband filter is expensive and difficult to popularize.Therefore,we selected the best band intervals for each year based on the high and low weight band intervals,and constructs the NSDI using the average reflectance of the best band intervals.The predicted R~2of the disease severity is up to 0.94.Since the optimal single wavelength and optimal band range in each year generally tend to close to the spectral range,located in 543nm-563nm(blue-green light range highly sensitive to plant pigments)and 653nm-666nm(red light range highly sensitive to plant diseases),they are set as the comprehensive band range.The NSDI was constructed based on the average reflectance of the comprehensive band interval,and the samples of 2018,2019,and 2020 were verified,and the R~2 was 0.95,0.91,and 0.94,respectively.In order to further verify the validity of the index,14 spectral indices commonly used in disease detection were applied to the same sample as a comparison,and the prediction effect was all lower than the NSDI presented in this study.Although there are many differences between samples in different years,the NSDI proposed in this study has universality and stability in the prediction of disease severity in three years,providing technical support for the development of multispectral cameras for rapid detection of FHB in the future.(2)For the harvested wheat grains,the PSR+3500 non-imaging spectrometer was used to collect spectral data.And a New Spectral Classification Index(NSCI)was designed by analyzing the spectral difference between healthy and diseased kernels in the 350-2500nm wavelength range.The index contains two parameters,reflectivity and first derivative,and was constructed in the form of normalized wavelength difference.The characteristic wavelength was calculated from the average spectral of the two classes of wheat,which are 1878nm and1887nm respectively.Frequency histograms were drawn for each class of index value,and gaussian curve fitting was performed for each histogram.Then,the intersection point of the Gaussian curves was used as the threshold to classify wheat kernels,with the classification accuracy of 0.97,sensitivity of 0.93,specificity of 0.99 and training time of 15.07 seconds.Meanwhile,the threshold could be tuned to adjust accuracy,sensitivity or specificity to satisfy different practical needs.In order to further verify the effectiveness of the proposed NSCI,three machine learning algorithms and four spectral indices for scab kernel classification were applied to the same sample for comparison.The results show that NSCI has high operating efficiency and low computational cost,and is more balanced in terms of classification efficiency and accuracy.We also applied the NSCI to kernel data in another year,and the classification results is promising.The proposed method has the potential for the rapid and simple detection of diseased kernel in wheat.