Performance Comparison of Adam and SGD Optimizers in Transfer Learning Based CNN for Banana Leaf Disease Classification

Banana leaf diseases significantly reduce crop productivity, yet automated detection systems based on deep learning often rely on limited datasets, where training stability and generalization become critical challenges. Although Convolutional Neural Networks (CNNs) have been widely applied for plant disease classification, systematic comparisons of optimization algorithms under small dataset conditions remain limited, particularly for banana leaf disease identification. This study addresses this gap by comparing the performance of Adaptive Moment Estimation (Adam) and Stochastic Gradient Descent (SGD) optimizers within a transfer learning–based CNN framework. Six pre-trained architectures VGG16, VGG19, ResNet50, DenseNet121, MobileNet, and NASNetMobile were evaluated using 1,652 annotated banana leaf images classified into Sigatoka, Cordana, Pestalotiopsis, and healthy leaves. Both optimizers were trained under identical experimental settings to ensure a fair comparison. Experimental results show that VGG19 achieved the highest accuracy, reaching 85% with Adam and 83% with SGD, while lightweight architecture exhibited lower performance due to underfitting. The findings demonstrate that optimizer selection plays a crucial role in improving CNN performance for banana leaf disease classification, especially when data availability is limited.