Interpretable Machine Learning Prediction of Bemisia tabaci Mortality Across Squash Cultivars and IPM Treatments

Management of the sweetpotato whitefly, Bemisia tabaci (Gennadius) (Hemiptera: Aleyrodidae), in squash is difficult because treatment efficacy may vary with host cultivar, treatment type, and time after application. This study developed an explainable machine learning framework to predict B. tabaci nymphal mortality and compare microbial, chemical, and combined IPM treatments under greenhouse conditions. Mortality was assessed at 3, 5, and 7 days post-inoculation (dpi) across eight squash cultivars treated with Beauveria bassiana GHA, Cordyceps javanica wfGA17, cyantraniliprole, imidacloprid, and selected fungal–insecticide combinations. Four regression models were compared using five-fold grouped cross-validation. Because the contemporaneous total nymph count was also the denominator used to calculate proportional mortality, it was excluded from the primary predictive models to avoid mathematical coupling with the response. Gradient Boosting showed the strongest predictive performance (R2 = 0.477, RMSE = 0.215, MAE = 0.156). Cultivar-averaged bootstrap predictions increased over time for all treatments. At 7 dpi, C. javanica + imidacloprid and B. bassiana + cyantraniliprole reached cultivar-averaged bootstrap mean predicted mortalities of 89.8% and 87.1%, respectively. SHAP analysis indicated that treatment identity and the temporal response structure represented by dpi and dpi2 were the strongest contributors to model predictions, whereas cultivar contributions were smaller. Sensitivity analysis showed that excluding total nymph count did not reduce Gradient Boosting performance. The framework provides a greenhouse-based predictive and interpretable approach for comparing IPM treatment responses, but independent field validation is required before operational use.