PERFORMANCE ANALYSIS OF FINE-TUNED EFFICIENTNET-B0 AND MOBILENETV2 FOR RICE LEAF DISEASE CLASSIFICATION WITH OPTIMIZER VARIATIONS ON MOBILE-BASED APPLICATION

Firmansyah, Egar (2026) PERFORMANCE ANALYSIS OF FINE-TUNED EFFICIENTNET-B0 AND MOBILENETV2 FOR RICE LEAF DISEASE CLASSIFICATION WITH OPTIMIZER VARIATIONS ON MOBILE-BASED APPLICATION. Undergraduate thesis, UPN Veteran Jawa Timur.

	Text (Cover) 22081010012.-cover.pdf Download (1MB)
	Text (Bab 1) 22081010012.-bab1.pdf Download (203kB)
	Text (Bab 2) 22081010012.-bab2.pdf Restricted to Repository staff only until 15 June 2029. Download (1MB)
	Text (Bab 3) 22081010012.-bab3.pdf Restricted to Repository staff only until 15 June 2029. Download (1MB)
	Text (Bab 4) 22081010012.-bab4.pdf Restricted to Repository staff only until 15 June 2029. Download (7MB)
	Text (Bab 5) 22081010012.-bab5.pdf Download (192kB)
	Text (Daftar pustaka) 22081010012.-daftarpustaka.pdf Download (198kB)
	Text (Lampiran) 22081010012.-lampiran.pdf Restricted to Repository staff only until 15 June 2029. Download (1MB)

Abstract

Rice leaf disease is one of the primary factors contributing to the decline of rice productivity in Indonesia. Accurate early detection is essential to enable farmers to take appropriate countermeasures before the disease spreads further. This study aims to analyze the performance of deep learning models based on Convolutional Neural Network (CNN) in classifying six classes of rice leaf diseases, namely Bacterial Leaf Blight, Brown Spot, Healthy Rice Leaf, Leaf Blast, Rice Hispa, and Sheath Blight, as well as implementing the best-performing model into a mobile application. Three model architectures were comparatively evaluated: MobileNetV2 with a frozen transfer learning approach, frozen EfficientNet-B0, and Fine-tuned EfficientNet-B0 which allows all base model parameters to be retrained. The search for optimal hyperparameter configurations was conducted through four sequential testing stages encompassing variations in batch size, epoch, learning rate, and optimizer (Adam, RMSprop, SGD) using a dataset of 4,770 images divided into training (3,333), validation (958), and testing (479) sets. The results demonstrate that the Fine-tuned EfficientNet-B0 model with a configuration of batch size 16, epoch 50, learning rate 0.001, and Adam optimizer achieved the best performance with a testing accuracy of 99.79% and a macro average f1-score of 0.9981, with only 1 misclassification out of 479 images. In comparison, MobileNetV2 achieved an accuracy of 92.48%, while frozen EfficientNet-B0 experienced a complete failure in the form of class collapse with an accuracy of only 19.62%. The best model was subsequently converted to TensorFlow Lite Standard format with a size of 16.56 MB (a 67.4% reduction from the original model) and integrated into the SIPADI mobile application (Sistem Pendeteksi Awal Penyakit Daun Padi) developed using the Flutter framework. Testing on a mobile device showed that the application is capable of performing accurate classification with an inference time of 198 ms, accompanied by educational information including disease descriptions, symptoms, treatment steps, and prevention measures for each disease class.

Item Type:

Thesis (Undergraduate)

Contributors:

Contribution	Contributors	NIDN/NIDK	Email
Thesis advisor	Muttaqin, Faisal	NIDN0030058602	faisalmuttaqin.if@upnjatim.ac.id
Thesis advisor	Nurlaili, Afina Lina	NIDN0013129303	afina.lina.if@upnjatim.ac.id

Subjects:

Q Science > QA Mathematics > QA76.87 Neural computers

Divisions:

Faculty of Computer Science > Departemen of Informatics

Depositing User:

EGAR EGAR FIRMANSYAH

Date Deposited:

15 Jun 2026 06:16

Last Modified:

15 Jun 2026 06:27

URI:

https://repository.upnjatim.ac.id/id/eprint/53983