A Comparative Study of Convolutional Neural Networks for Oregano Diseases Classification in A Novel Dataset

Peru is South America's largest producer and ex-porter of oregano (Origanum vulgare L.), with the Tacna region being the most significant contributor, cultivating 2,499 hectares and producing 11,946 tons, accounting for 65% of the national production. To mitigate the impact of oregano diseases, which lead to substantial crop losses and negatively affect the economy, it is crucial to develop and upgrade technologies for disease management. Accurate classification of oregano diseases can prevent quality deterioration and yield reductions. However, the availability of data and the use of advanced technologies in oregano cultivation are still in their early stages. This study addresses these shortcomings by developing and publishing a novel public dataset for the classification of the most common oregano diseases-Alternaria, Powdery Mildew (Oidium), and Mosaic-using crops from Susapaya - Tarata in Tacna, Peru. To classify oregano diseases, we conducted a comprehensive comparative study, applying Transfer Learning and Fine-Tuning techniques to evaluate the performance of VGG16, MobileNet, MobileNetV2, and NasNetLarge architectures with modifications in the classifier. The experiments revealed that: (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$i$</tex>) Fine-Tuning improved validation set accuracy by 4.47% compared to Transfer Learning, (ii) the VGG16 architecture achieved the highest accuracy at 0.993921, (iii) MobileNet and MobileNetV2 models performed similarly, (iv) the NASNetLarge model showed poor performance when only Transfer Learning was applied, and (v) there was a strong significant relationship <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(R^{2}=\mathbf{0.84})$</tex> between the number of parameters and size (in MB) with the performance of the architectures on our dataset. We ensure the reproducibility of the results by making the novel dataset and the best model publicly available on GitHub.