DETEKSI PENYAKIT DAUN TEH MENGGUNAKAN TRANSFER LEARNING MODEL EFFICIENTNETV2B3 BERBASIS MOBILE

Tea Leaf Disease Detection Using Transfer Learning with the EfficientNetV2B3 Model Based on Mobile

Indonesia merupakan salah satu produsen dan eksportir teh terbesar di dunia. Namun, produktivitas teh menurun akibat ancaman perubahan iklim yang menyebabkan munculnya berbagai hama serta meningkatnya kerentanan terhadap penyakit, sehingga memengaruhi kualitas dan kesehatan tanaman teh. Identifikasi manual yang subjektif sering menyebabkan keterlambatan penanganan. Untuk mengatasi permasalahan tersebut, penelitian ini mengembangkan sistem deteksi penyakit daun teh berbasis mobile menggunakan Transfer Learning model EfficientNetV2B3. Dataset yang dipakai adalah TeaLeafBD dari Kaggle berisi 5.276 citra daun teh yang terdiri dari 7 kelas. Optimizer SGD, RMSProp, dan Adam dibandingkan untuk mendapatkan performa terbaik dengan metrik evaluasi accuracy, precision, recall, dan F1-score. Hasil penelitian menunjukkan bahwa RMSProp memiliki performa terbaik dengan akurasi validasi 96,30%. Model terbaik kemudian diintegrasikan ke dalam aplikasi mobile menggunakan framework Flutter yang mampu menerima input citra dari kamera maupun galeri, menampilkan hasil prediksi beserta nama penyakit, serta menyimpan riwayat deteksi secara lokal menggunakan SQLite tanpa memerlukan koneksi internet. Sistem ini diharapkan dapat membantu petani teh melakukan deteksi penyakit secara mandiri dan cepat, sehingga mendukung peningkatan efisiensi, produktivitas, dan kualitas hasil produksi teh di Indonesia.

Indonesia is one of the world's largest tea producers and exporters. However, tea productivity has been declining due to climate change, which has led to the emergence of various pests and increased susceptibility to diseases, affecting the quality and health of tea plants. Subjective manual identification often causes delays in treatment. To address this problem, this study developed a mobile-based tea leaf disease detection system using Transfer Learning with the EfficientNetV2B3 model. The dataset used is TeaLeafBD from Kaggle, containing 5,276 tea leaf images across 7 classes. The SGD, RMSProp, and Adam optimizers were compared to obtain the best performance using accuracy, precision, recall, and F1-score as evaluation metrics. The results show that RMSProp achieved the best performance with a validation accuracy of 96.30%. The best model was then integrated into a mobile application using the Flutter framework, capable of receiving image input from the camera or gallery, displaying prediction results along with the disease name, and storing detection history locally using SQLite without requiring an internet connection. This system is expected to help tea farmers detect diseases independently and quickly, thereby supporting improved efficiency, productivity, and the quality of tea production in Indonesia.