Machine Learning-Assisted Experimental Electronic Nose Measurements for Biopesticide Treatment Discrimination in Tomatoes

This study evaluated whether a compact electronic nose could distinguish intact tomatoes treated with wood vinegar or neem oil from those without experimental biopesticide application. The dataset comprised 38 tomatoes and 228 technical runs, including six additional controls measured in triplicate initially and after 16 h: three untreated and three water-sprayed. Six commercial SnO2 sensors recorded voltages during a 180 s cycle. Six classifiers were compared using identical tomato-wise nested cross-validation with three outer folds, two inner folds, and five repetitions. With 54 full-cycle features, linear discriminant analysis (LDA) achieved 97.02% accuracy and 95.72% balanced accuracy; an artificial neural network (ANN) achieved 96.93% and 95.14%, respectively. Paired conditional intervals did not establish superiority of either model. Models trained on the original 32 fruits produced false biopesticide alarms in the new controls, including 17/36 ANN predictions. Incorporating eligible control fruit into training improved matched held-out performance, mainly under the new negative conditions. The separate two-treatment task included 32 fruits. These findings support exploratory recognition of patterns associated with application and show the value of procedural controls. Small control groups, unlogged chamber conditions, and missing chemical reference measurements limit chemical attribution and external screening claims.