Varroa destructor is the most devastating ectoparasite of Apis mellifera, and early detection is critical for colony survival. This study systematically investigated how image preprocessing, model architecture, and feature map resolution jointly affect classification accuracy and Grad-CAM++ explainability in deep-learning-based Varroa detection. From comb-surface images of 20 A. mellifera colonies, 3400 region-of-interest images were processed through 12 preprocessing pipelines combining deblurring, histogram normalization, morphology-preserving resizing, and non-morphological resizing. Nineteen CNN architectures, including VarroaNet — a custom lightweight model with configurable channel attention — were screened across all pipelines, and the top six further evaluated at four feature-map resolutions (7 × 7 to 56 × 56); the two stages together comprised 1,548 classification training runs across 516 configurations. Resizing consistently improved classification accuracy, whereas histogram normalization degraded it. VarroaNet (r = 8) achieved the highest mean accuracy across configurations (97.28%) with the lowest cross-configuration variability (CV = 1.47%). The 28 × 28 resolution was jointly optimal for classification and localization at minimal computational overhead, whereas 56 × 56 degraded performance. Notably, classification accuracy and localization quality did not always coincide—the highest-accuracy configuration (ShuffleNet-V2-x1.0 at 14 × 14, 97.34%) achieved an IoU@30 of only 0.160, underscoring the need for explicit localization evaluation. Morphology-preserving resizing achieved higher localization efficiency with zero morphological distortion. The recommended configuration—VarroaNet (r = 8) at 28 × 28 with deblurred MR preprocessing—achieved the highest localization performance (Pointing Game = 0.927), indicating correct attention to the mite region in 92.7% of infested test images.
Lee et al. (Sun,) studied this question.