<p>Diseases such as early and late blight, which affect potato leaves, are serious hazards to global agriculture and food supply. Efficient management of crops relies on the detection of these diseases in an accurate and efficient manner. In order to identify potato leaf illness, this study involves 1500 images of early, late, and healthy potato leaves. The article compares different models such as SVM, KNN, decision tree, random forest, and CNN, with statistical testing, and found out that CNN has better accuracy as compared to other methods. Afterwards, the proposed method takes advantage of CNN to recognize complicated pattern dependencies from images, and PUMA guarantees the optimal management of hyperparameters. When CNN is integrated with the PUMA optimizer, the proposed CNN-PUMA architecture enhances performance in the detection of disease. The experimental results confirmed that the CNN-PUMA model was superior to the other optimizers such as PSO, GA, GGO, and GWO in terms of the metrics like accuracy, sensitivity (recall), specificity, positive predictive value (PPV), and F1-score. The testing results further indicate significant improvement, with precision scoring 98.2 and recall at 97.5, with accuracy at 99.1%.</p>