Deep Learning Approach for Oil Palm Leaf Disease Classification Using VGG16 Enhanced with Adam Optimization

Abstract

Early detection of oil palm leaf diseases is essential to minimize economic losses and ensure sustainable plantation management. This study proposes a deep learning approach using the VGG16 architecture optimized with the Adam algorithm to classify oil palm leaves into three categories: healthy, infected, and initial infection. The dataset was obtained from Roboflow and preprocessed through cropping, annotation, and standardization before being split into training, validation, and testing sets. Experimental evaluations were performed across multiple training epochs and compared against two baseline models: a shallow Convolutional Neural Network (CNN) and YOLOv11-CLS (version S). Results show that VGG16 combined with Adam achieved the highest accuracy of 97% at 25 and 50 epochs, with balanced precision and recall across all classes. In contrast, the baseline CNN reached a maximum accuracy of 88%, while YOLOv11-CLS produced fluctuating results with a peak accuracy of 82%. Statistical significance testing confirmed that the performance improvements of VGG16 + Adam were consistent and reliable, validating its suitability for practical implementation in precision agriculture. These findings highlight the potential of combining deep architectures with adaptive optimization to enhance disease diagnosis in oil palm plantations and reduce reliance on manual inspections.