Pengembangan Website Deteksi dan Visualisasi Kesehatan Tanaman Cabai Berbasis Citra Daun Menggunakan Model EfficientNet-B0

Abstract

Identifikasi penyakit tanaman cabai di Balai Besar Pengembangan dan Penerapan Modernisasi Pertanian masih dilakukan secara manual melalui pengamatan visual, sehingga membutuhkan waktu lebih lama dan berpotensi menimbulkan kesalahan diagnosis. Penelitian ini bertujuan mengembangkan sistem informasi berbasis web yang mengintegrasikan model EfficientNet-B0 untuk mendeteksi kondisi kesehatan tanaman cabai berbasis citra daun serta menyajikan visualisasi hasil deteksi. Pengembangan sistem menggunakan metodologi Agile dan pendekatan transfer learning pada 450 citra daun cabai yang terdiri atas tiga kelas, yaitu sehat, terserang hama, dan terinfeksi virus yellow leaf curl. Evaluasi dilakukan melalui pengujian performa model, black-box testing, dan User Acceptance Test (UAT) bersama penyuluh pertanian. Sistem yang dihasilkan memiliki fitur unggah citra, prediksi kondisi daun, penyimpanan riwayat deteksi, visualisasi hasil deteksi, manajemen pengguna, serta pengelolaan model melalui antarmuka admin. Hasil black-box testing menunjukkan seluruh fitur utama berjalan sesuai skenario, dan hasil UAT menunjukkan sistem diterima pengguna. Model memperoleh akurasi 100% pada data validasi, tetapi hanya 36% pada data uji eksternal. Kesenjangan ini menunjukkan keterbatasan generalisasi model, sehingga sistem lebih tepat digunakan sebagai prototipe atau alat bantu awal. Pengembangan selanjutnya perlu difokuskan pada perluasan variasi dataset dan optimalisasi model.

The identification of chili plant diseases at the Center for the Development and Application of Agricultural Modernization is still carried out manually through visual observation, which is time-consuming and may lead to misdiagnosis. This study aims to develop a web-based information system that integrates an EfficientNet-B0 model to detect chili plant health conditions from leaf images and visualize the detection results. The system was developed using Agile methodology and a transfer learning approach on 450 chili leaf images consisting of three classes: healthy leaves, pest-infected leaves, and yellow leaf curl virus-infected leaves. Evaluation was conducted through model performance testing, black-box testing, and User Acceptance Test (UAT) with agricultural extension workers. The system provides image upload, leaf condition prediction, detection history storage, result visualization, user management, and model management through an admin interface. The black-box testing results showed that all main features functioned as expected, and the UAT results showed that the system was accepted. The model achieved 100% accuracy on validation data but only 36% accuracy on external test data. This gap indicates limited model generalization, suggesting that the system is currently more suitable as a prototype, while future development should focus on expanding dataset diversity and optimizing the model.