Electronic Theses and Dissertation

Tanaman cabai merupakan komoditas hortikultura penting yang produktivitasnya dapat menurun akibat serangan penyakit pada daun. Identifikasi penyakit secara manual memiliki keterbatasan karena bergantung pada pengalaman pengamat dan dapat dipengaruhi oleh kemiripan gejala antar kelas. Penelitian ini bertujuan mengembangkan model klasifikasi penyakit daun cabai menggunakan citra visual RGB dan citra termal color mapping dengan pendekatan multimodal dan semi-supervised learning. Dataset yang digunakan terdiri dari 400 pasangan citra RGB-termal pada empat kelas, yaitu Bercak, Keriting, Kuning, dan Sehat. Dataset dibagi menjadi data latih, validasi, uji, dan data tanpa label untuk mendukung skenario semi-supervised learning.
Tahapan penelitian meliputi pemilihan backbone CNN menggunakan EfficientNet-B0, ResNet50, dan DenseNet121; hyperparameter tuning; pengembangan model RGB supervised; penerapan FixMatch untuk RGB semi-supervised; pengembangan model multimodal supervised berbasis dua cabang DenseNet121; serta pengembangan model multimodal semi-supervised menggunakan FixMatch. Hasil eksperimen menunjukkan bahwa DenseNet121 merupakan backbone terbaik dengan akurasi 90,83%. Pada evaluasi akhir, model RGB supervised memperoleh akurasi 94,16%, RGB semi-supervised memperoleh 95,00%, multimodal supervised memperoleh 95,83%, dan multimodal semi-supervised memperoleh 96,66%. Hasil tersebut menunjukkan bahwa citra termal memberikan informasi komplementer terhadap citra RGB, sedangkan data tanpa label melalui pseudo-labeling dan consistency regularization dapat meningkatkan kemampuan generalisasi model. Model terbaik adalah multimodal semi-supervised berbasis DenseNet121 dengan accuracy precision, recall, dan F1-score sekitar 0,97 pada data uji.

Chili is an important horticultural commodity whose productivity can be reduced by leaf diseases. Manual disease identification is often subjective and may be challenging when symptoms among disease classes are visually similar. This study aims to develop a chili leaf disease classification model using RGB visual images and thermal color-mapping images through multimodal and semi-supervised learning. The dataset consists of 400 paired RGB-thermal samples from four classes: leaf spot, leaf curl, yellow leaf, and healthy. The dataset is divided into training, validation, testing, and unlabeled subsets to support the semi-supervised learning scenario. The research workflow includes CNN backbone selection using EfficientNet-B0, ResNet50, and DenseNet121; hyperparameter tuning; RGB supervised learning; RGB semi-supervised learning using FixMatch; supervised multimodal learning with a dual-branch DenseNet121 architecture; and multimodal semi-supervised learning using FixMatch. The experimental results show that DenseNet121 is the best backbone with an accuracy of 90.83%. In the final evaluation, the RGB supervised model achieves 94.16% accuracy, the RGB semi-supervised model achieves 95.00%, the supervised multimodal model achieves 95.83%, and the multimodal semi-supervised model achieves 96.66%. These results indicate that thermal images provide complementary information to RGB images, while unlabeled data utilization through pseudo-labeling and consistency regularization improves model generalization. The best model is the DenseNet121-based multimodal semi-supervised model, achieving approximately 0.97 accuracy, precision, recall, and F1-score on the test set.