Recent research in plant disease classification leverages advanced deep learning models and both public and author-generated datasets. The study adopts a two-step approach: initially conducting a comparative analysis to identify the most effective convolutional neural network (CNN) among established architectures and custom models, and subsequently employing deep learning optimizers to enhance the performance of the top-performing model. Various metrics, including validation accuracy/loss, F1-score, precision, recall, and training epochs, are used to evaluate popular deep learning architectures such as VGG16, VGG19, ResNet-50, and ResNet-101. All architectures are trained on the extensive PlantVillage dataset, containing 70,295 images of diseases across 14 plant species. In a study focused on disease prediction using Keras with TensorFlow, the findings highlight that the RESNET101 model, trained with the Adam optimizer, achieved the highest macro and micro F1-scores, reaching 0.9417 and 0.9425, along with exceptional testing and validation accuracy rates of 94.25 and 93.91%. These outcomes signify a substantial improvement over earlier models, with the novelty of this study lying in the class-wise comparison of metrics, underscoring the innovative nature of this research. Furthermore, the methodology holds significant potential for broader applications in the field of agriculture, as early and accurate disease detection plays a pivotal role in crop health management, leading to increased yields, reduced pesticide use, and overall sustainable agriculture practices. Class-wise metrics, including precision, recall, and F1 scores, further contribute to a comprehensive evaluation of disease detection performance, enhancing its practical utility.