AJI, HILARIUS SETA SAPTA (2026) PERFORMANCE COMPARISON OF EFFICIENTNETV2, MOBILENETV3, AND THEIR INTEGRATION WITH CAPSULE NETWORK FOR IMAGEBASED RICE PLANT DISEASE CLASSIFICATION. S1 thesis, UNIVERSITAS KATOLIK SOEGIJAPRANATA SEMARANG.

Preview	Text 22.K3.0009 - HILARIUS SETA SAPTA AJI-COVER_a.pdf Download (1MB) | Preview
	Text 22.K3.0009 - HILARIUS SETA SAPTA AJI-ISII_a.pdf Restricted to Registered users only Download (6MB)
Preview	Text 22.K3.0009 - HILARIUS SETA SAPTA AJI-DAPUS_a.pdf Download (995kB) | Preview
	Text 22.K3.0009 - HILARIUS SETA SAPTA AJI-LAMP_a.pdf Restricted to Registered users only Download (5MB)

Abstract

Rice productivity in Indonesia is often disrupted by leaf and neck diseases such as bacterial leaf blight, brown spot, and leaf blast, while manual detection remains inefficient and error-prone. This study compares the performance of EfficientNetV2B0 and MobileNetV3, both individually and in hybrid combination with Capsule Neural Network (CapsNet), for image-based classification of ten rice plant disease classes. A public dataset of approximately 30,235 images was used, with all models trained through transfer learning under both augmented and nonaugmented conditions, and evaluated using accuracy, precision, recall, F1-score, confusion matrix, inference time, and model size. Experimental results show that all six model configurations achieved excellent performance, with accuracy exceeding 99.5%. The hybrid EfficientNetV2B0+CapsNet model achieved the highest accuracy and F1-score of 99.784%, while MobileNetV3 reached its best accuracy of 99.736%. Data augmentation was found to provide a modest performance improvement compared to training without augmentation. However, integrating CapsNet increased training time and model complexity without producing a proportional gain in classification performance, making EfficientNetV2B0+CapsNet less efficient for real-world deployment despite its highest accuracy. In contrast, MobileNetV3 offered the smallest model size and fastest inference time, positioning it as the most suitable option for resource-constrained or edge-device implementation. These findings support the development of an accurate, efficient, and automated rice disease detection system for precision agriculture.

Actions (login required)

View Item