Exploring ensemble learning methods with labor-free line-by-line radiometric calibration for classifying the damage types of rice planthopper from hyperspectral images

The white-backed and brown planthopper are major rice pests whose co-infestation severely reduce growth and yield. Thus, timely and accurately identification of infested types is essential for precise pesticide application and effective pest management. This study assessed ensemble learning methods for classifying planthopper infestation types using hyperspectral images and explored their mechanisms through controlled indoor experiments. A novel labor-free, line-by-line radiometric correction method was developed to migrate distortions caused by uneven illumination during hyperspectral image acquisition. The effects of different correction strategies on infestation-type classification were evaluated using ensemble learning models. Additionally, SHAP-derived feature contributions were exploited to identify key factors and clarify the underlying mechanisms. The results showed that the proposed line-by-line radiometric correction significantly improved the consistency of rice hyperspectral data by reducing illumination-induced distortions. Based on the conventionally corrected data, XGBoost and Stacking method achieved superior performance in identifying planthopper damage types, with OA and Kappa exceeding 0.564 and 0.344, respectively. Their performance improved by more than 20% using line-by-line correction, with the highest OA and Kappa reaching 0.682 and 0.551. The blue (450-520 nm) and near-infrared (750-900 nm) regions contributed most to the task.