Early Detection of Powdery Mildew Disease and Accurate Quantification of Its Severity Using Hyperspectral Images in Wheat

Early detection of the crop disease using agricultural remote sensing is crucial as a precaution against its spread. However, the traditional method, relying on the disease symptoms, is lagging. Here, an early detection model using machine learning with hyperspectral images is presented. This study first extracted the normalized difference texture indices (NDTIs) and vegetation indices (VIs) to enhance the difference between healthy and powdery mildew wheat. Then, a partial least-squares linear discrimination analysis was applied to detect powdery mildew with the combined optimal features (i.e., VIs & NDTIs). Further, a regression model on the partial least-squares regression was developed to estimate disease severity (DS). The results show that the discriminant model with the combined VIs & NDTIs improved the ability for early identification of the infected leaves, with an overall accuracy value and Kappa coefficient over 82.35% and 0.56 respectively, and with inconspicuous symptoms which were difficult to identify as symptoms of the disease using the traditional method. Furthermore, the calibrated and validated DS estimation model reached good performance as the coefficient of determination (R-2) was over 0.748 and 0.722, respectively. Therefore, this methodology for detection, as well as the quantification model, is promising for early disease detection in crops.

Automated summary

This study introduced an early detection model for crop diseases using hyperspectral images and machine learning, which enhanced the accuracy of early identification of infected leaves by combining VIs and NDTIs features, as well as developed a partial least-squares regression model for estimating disease severity. The results showed promising ability for early disease detection and quantification in crops, with high overall accuracy and coefficient values.