Early and accurate identification of crop diseases is essential for minimizing agricultural losses and improving food security. In this study, we present a comparative analysis of five widely used CNN architectures (AlexNet, VGGNet, GoogleNet, ResNet, and DenseNet) for the classification of potato leaf diseases, including healthy, early blight and late blight classes. Each model was trained and evaluated on the same curated dataset under identical experimental conditions. Furthermore, a soft voting ensemble method was employed to combine the predictive strengths of the individual models. Performance was assessed using confusion matrices, standard classification metrics (accuracy, precision, recall, F1-score), learning curves, and AUC-ROC analysis. The results show that while GoogleNet and DenseNet achieved the highest individual accuracies, the ensemble model significantly outperformed all standalone architectures with an accuracy of 98.90% and perfect AUC scores across all classes. These findings demonstrate the potential of ensemble DL for robust and scalable crop disease detection in precision agriculture.