Effective plant disease diagnosis is key to sustainable farming, but data scarcity remains a significant hurdle. This research examined machine learning models trained on very small datasets enhanced by varied image processing techniques to identify the most reliable approach. The Random Walk Segmented dataset initially seemed more promising, excelling in both precision and accuracy. However, its performance faltered with a random-like ROC-AUC score, suggesting unreliability. In contrast, the Graph Cuts Segmented dataset, despite trailing in precision and accuracy, demonstrated greater consistency with a higher ROC-AUC score. These results highlight the critical need to use diverse metrics for evaluating machine learning models, emphasising that reliability cannot rely solely on accuracy. The study sheds light on advancing plant disease diagnosis in environments constrained by limited data, paving the way for more robust solutions tailored to resource-scarce contexts.
Enow et al. (Thu,) studied this question.