Classification of Cassava Leaf Diseases Using ResNet50 CNN Architecture Based on Digital Images

Abstract

Cassava (Manihot esculenta) is an important agricultural commodity in Indonesia, but its productivity can decline due to leaf diseases such as Cassava Mosaic Disease (CMD), Cassava Green Mottle (CGM), Cassava Bacterial Blight (CBB), and Cassava Brown Streak Disease (CBSD). These four diseases exhibit overlapping visual symptoms such as chlorosis, spots, and leaf discoloration, making them difficult to distinguish manually. This study aims to create a digital- based cassava leaf image classification system using the Convolutional Neural Network (CNN) algorithm and ResNet50 architecture. The dataset used consists of 9,436 cassava leaf images taken from the TensorFlow platform and processed through resizing, normalization, selective augmentation, and the application of transfer learning. The experiment compared various optimizer configurations, learning rates, batch sizes, and balanced and unbalanced dataset scenarios. The evaluation was conducted using accuracy, precision, recall, and F1-score. The results show that the best performance was obtained on an unbalanced dataset using the Adam optimizer (learning rate 0.001; batch size 64) with an accuracy of 80.69% and an F1-score of 79.76%. Meanwhile, balancing the dataset actually reduced performance to an accuracy of 77.14% and an F1-score of 76.48%. Analysis of the loss curve and confusion matrix confirmed that the natural data distribution provided more stable generalization, although misclassification still occurred in classes with similar visual symptoms. These findings indicate that ResNet50 is effective for classifying cassava leaf diseases and has the potential to support early detection in digital agriculture practices.