Detection and discrimination of pests and diseases in winter wheat based on spectral indices and kernel discriminant analysis

Abstract

Kernel discriminant analysis (KDA) can be used as a feasible strategy for identifying plant stresses, especially for detection of pests and diseases, considering the highly nonlinear distribution of hyperspectral absorption features that respond to biophysical variations in plants caused by foliar lesions. However, traditional computation of the kernel projection features produced by hyperspectral data is always affected by redundant information among the numerous wavelengths, subsequently leading to dimension disaster. In order to alleviate this problem, the aim of this study is to propose a spectral vegetation indices-based kernel discriminant approach (SVIKDA) for the detection and classification of yellow rust, aphid, and powdery mildew in winter wheat at the leaf and canopy level. Leaf and canopy level hyperspectral reflectance datasets were measured with a total of 314 and 187 samples, respectively. Fourteen Spectral Vegetation Indices (SVIs) related to foliar biophysical variations were employed as the input sample space; then, by using correlation analysis and independent t-tests, redundant information among SVIs was removed. Subsequently, a Gaussian kernel function was utilized to cast discriminant analysis into a nonlinear framework. Finally, using 5-fold cross validation, performance and transferability of this approach were evaluated. Our results revealed that the SVIKDA outperformed conventional linear discriminant approach on detection and classification among healthy wheat leaves and leaves infected with yellow rust, aphids, and powdery mildew. At the leaf level, the classification returned the overall accuracies (OA) of 82.9%, 89.2%, 87.9% for three occurrence levels, i.e. slight, moderate, and severe (Kappa>0.85). Depending on the types and severities of infestations, the classification accuracy was between 76% and 95%; At the canopy level, the multiple classifications between healthy leaves and leaves with damages from the three different infestations still achieved an accuracy greater than 87% (Kappa=0.84). In addition, this approach was also successfully applied in disease index (DI) estimation for a specific infestation at the leaf level, and optimal DI estimation returned high coefficients of determination (R2>0.7). Furthermore, compared with the commonly used automatic classification algorithm, the SVIKDA achieved an accurate classification without losing the pathological basis of input variables. The results suggest that this method has reliable transferability and great robustness in detecting and discriminating pests and diseases for guiding precision plant protection.