Identification of wheat powdery mildew using in-situ hyperspectral data and linear regression and support vector machines

To fully understand the spectral response characteristics of powdery mildew (PM) on winter wheat, in-situ hyperspectral data were collected and comparatively analyzed. The center distance method was first used to remove the abnormal spectral response bands using the red-edge position. Subsequently, the Relief-F algorithm and correlation analysis were jointly introduced to identify the best bands sensitive to the PM. The 636 nm in the visible region and the 784 nm in the near-infrared region were finally assured to develop a new vegetation index (NDVI1) according to the generation mechanism of normalized difference vegetation index (NDVI). Besides, a total of ten other vegetation indices commonly used in previous studies were calculated for comparatively evaluating the performance of NDVI1. Two types of sample data (only diseased samples, and both diseased and healthy samples) and three classifiers were comparatively used to estimate the disease, including a linear regression, support vector machine (SVM) and least squares support vector machine (LS-SVM) models. The results show that the linear regression model based on the NDVI1 except for the Modified Simple Ratio (MSR) is generally the best for the two sample types, giving a coefficient of determination (R-2) of 0.75 and 0.49, respectively. Conversely, the SVM and LS-SVM models provide the best estimation accuracy using the K-fold cross-validation. In general, the overall classification accuracy of the SVM is higher than that of the LS-SVM, but the LS-SVM is more efficient regarding running time. The results of this study can provide a useful guideline for wheat PM estimation using the ground-based hyperspectral data.