This work compares the performances of a custom CNN, ResNet50, DenseNet121, and InceptionV3 for the classification of sugarcane diseases. The ResNet50 had the highest validation accuracy of 80.16% but with the longest training time of 883.07 s; thus, it becomes very suitable for applications where high accuracy is the focus. A Custom CNN had a reasonable balance in terms of accuracy of 74.01% in a relatively fast training time of 52.61 s, hence suitable for applications when the time available is scanty. The performance of DenseNet121 was average, whereas the performance of InceptionV3 was poor by far. The results indicate a trade-off between the solution space of high accuracy and computational efficiency. Further research will concentrate on data augmentation, lightweight models, and real-time deployment for improvements in practicality regarding agricultural applications.