Integrating feature selection and explainable CNN for identification and classification of pests and beneficial insects.

Reliable identification of agricultural pests and beneficial insects is crucial for sustainable crop protection and ecological balance, yet most vision-based models remain black boxes and require high-dimensional features. This paper proposes an explainable hybrid insect-classification framework that combines convolutional neural network (CNN) feature extraction with a dual-XAI feature selection strategy. SHapley Additive exPlanations (SHAP) and Permutation Feature Importance (PFI) are applied in parallel to rank handcrafted and CNN-derived features, and their intersection yields a compact, biologically meaningful subset for final classification. The selected features are evaluated using lightweight classifiers and a hybrid ensemble, enabling accurate inference under field variability. Experiments on a curated, balanced dataset of four classes (Colorado potato beetle, green peach aphid, seven-spot ladybird, and healthy leaves) collected under diverse lighting and background conditions achieve 96.7% overall accuracy, with precision, recall, and F1-scores all above 96%. Importantly, performance remains stable when reducing dimensionality, retaining ≥90% accuracy using only the top 11 hybrid-selected features. These results demonstrate that integrating SHAP and PFI improves both robustness and interpretability, supporting practical deployment for automated pest monitoring and precision agriculture.