Wheat Fusarium Head Blight Identification Based on New Index Using in Situ Hyperspectral Data

Wheat Fusarium head blight (FHB), a serious disease caused by the fungus Gibberella zeae, occurs in many parts of the world. Wheat infected with FHB can turn yellow and dry ears, shrivel grains and produce a toxin called deoxynivalenol (DON), which seriously affects wheat quality and yield. Accurate identification of infection severity of wheat under FHB stress is very important for crop growth diagnosis and disease precision management. Wheat FHB is easy to develop, accompanied by an insignificant early symptoms and rapid progression, so hyperspectral data with abundant band information have significant advantages in wheat disease identification. The aim of this study was to accurately identify the degree of wheat FHB infection by constructing new spectral indices based on hyperspectral data. Multiplicative Scatter Correction (MSC) method was first used to preprocess the original hyperspectral data to reduce the spectral differences caused by different wheat varieties and geographical environment. Then, we performed first derivative on the hyperspectral data to eliminate noise and enhance the weak spectral signal. On the basis of comprehensive analysis of hyperspectral data, three new spectral indices mainly include Normalized difference spectral index (DNDSI), Ratio spectral index (DRSI) and Differential spectral index (DDSI) were constructed. All wavelengths in the region of 350-900nm were considered to construct new indices to provide an accurate and robust FHB identification. Linear regression methods were adopted to assess the ability of DNDSI, DRSI, and DDSI to identify FHB, and were compared with traditional vegetation indices. The identification capability of Multivariable regression methods, such as random forest regression (RFR) using the whole hyperspectral data and subwindow permutation analysis fusion RFR (SPA-RFR) using selected bands, were also used for testing. The results showed that compared with the traditional vegetation indices, the new indices constructed in our study showed better ability in the identification of wheat FHB. DRSI (501, 682) using the first derivative values at 501nm and 682nm, was found to be the most accurate and reliable index for the disease identification of wheat infection ($R^{2}$ = 0.801). The prediction ability of the index exceeded that of the RF model using the whole bands ($R^{2}$ = 0.738), and was comparable to that of the SPA-RFR model using the selected bands ($R^{2}$ = 0.795). The new method proposed in this study can greatly improve the identification level of wheat-infected FHB, and can also provide valuable theoretical and method reference for other crop diseases diagnosis and accurate identification based on hyperspectral technology.