Computer vision-aided classification of insecticide-induced behavioral patterns in Aphis gossypii (Hemiptera: Aphididae).

Electrophysiological approaches are widely used to characterize insecticide target sites, whereas behavioral symptoms associated with different modes of action remain less quantitatively explored. Here, we developed a computer vision-based framework to quantify insecticide-induced behavioral responses in the cotton aphid, Aphis gossypii Glover (Hemiptera: Aphididae). Convolutional neural networks were trained on 253 video recordings of aphids treated with neonicotinoid, organophosphate, or pyrethroid insecticides. A top-down pose-estimation approach enabled reliable detection of 11 body parts, from which 66 behavioral features were extracted. Feature selection and machine-learning classification were used to evaluate differentiation among insecticide classes. Among 31 tested algorithms, bagging ensemble trees showed the highest performance, achieving 74.36% accuracy and a macro-averaged Area Under the Curve (AUC) of 0.895 on the independent test set. Classification of additional compounds within trained classes showed moderate predictive consistency, whereas compounds with untrained modes of action produced dispersed prediction patterns across classes. These findings demonstrate that automated behavioral profiling can capture class-associated patterns under controlled conditions and may provide a quantitative complement to conventional toxicological assays.