Early detection of wheat powdery mildew: A multi-source in situ remote sensing approach enabled by stacked ensemble learning

Powdery mildew seriously hinders photosynthesis and nutrient accumulation in wheat, and its early detection holds the key to enhancing control efficacy. In this research, solar-induced chlorophyll fluorescence (SIF) parameters were derived from radiance and reflectance data, while vegetation indices (VI) were computed using reflectance. A suite of feature selection methods, including shadow feature (Boruta), feature selection (ReliefF), minimum redundancy maximum correlation (mRMR), and random forest (RF). Models were developed on the back propagation (BP) neural network, support vector regression (SVR), and partial least squares regression (PLSR). Furthermore, a stacking ensemble strategy was adopted, utilizing RF and decision tree (DT) algorithms as meta-models to integrate the predictions from base models. The findings revealed that the Boruta method selected a well-balanced number of feature parameters with normalized weights. The multi-source model (SIF + VI) is superior to the single-source model (SIF or VI). The BP model exhibited high accuracy in wheat disease monitoring, particularly during the initial infection phases. The multi-regressor stacked with RF ensemble model (MRSRF) generally surpassed the multi-regressor stacked with DT ensemble model (MRSDT), especially in the initial infection stage, where the MRSRF model's average R² was 13.03 % higher than that of the BP model. To validate these conclusions, reflectance data simulated by the PROSAIL model (PROSPECT and SAIL) were utilized. The Boruta-MRSRF model demonstrated exceptional advantages in early detection, achieving an R² greater than 0.90 at all infection stages. This study provides effective ideas and methods for the active prevention and control of crop diseases, which are of great significance for ensuring agricultural production.