Electronic Theses and Dissertation

Penelitian ini bertujuan untuk mengklasifikasikan gambar tanaman herbal di Aceh
dengan menggunakan metode deep learning berbasis arsitektur ResNet50V2. Dataset
yang digunakan terdiri dari 32 kelas tanaman herbal dengan total 1.044 gambar yang
diperoleh melalui pengambilan gambar langsung dan sumber online. Data dibagi
menjadi dataset train dan dataset test dengan rasio 80:20, kemudian dilakukan 7 jenis
augmentasi secara offline untuk meningkatkan variasi data, diantaranya rotasi,
flipping, translasi, dan scaling. Proses pelatihan dilakukan tanpa augmentasi real-
time, diawali dengan baseline model (tanpa augmentasi) untuk mendapatkan performa
awal. Selanjutnya, dilakukan sepuluh percobaan dengan strategi fine-tuning berbeda,
mulai dari freeze seluruh layer hingga membuka 70 layer terakhir. Tahap pelatihan
menggunakan callback seperti EarlyStopping, ModelCheckpoint, dan
ReduceLROnPlateau untuk menghindari overfitting dan mengoptimalkan performa.
Hasil pengujian menunjukkan bahwa baseline model tanpa augmentasi menghasilkan
akurasi 73,89%, sedangkan model terbaik diperoleh pada percobaan kesepuluh dengan
strategi fine-tuning 70 layer terakhir, penambahan dense layer (256 dan 128 neuron)
serta dua dropout (masing-masing 0,5). Model ini mencapai akurasi tertinggi sebesar
84,07% pada data uji. Temuan ini membuktikan bahwa augmentasi data dan fine-
tuning terstruktur mampu meningkatkan performa klasifikasi citra tanaman herbal di
Aceh secara signifikan.
Kata kunci: ResNet50V2, fine-tuning, augmentasi data, tanaman herbal aceh,
klasifikasi gambar

This research aims to classify images of traditional Acehnese herbal plants using a deep learning approach based on the ResNet50V2 architecture. The dataset consists of 32 classes of herbal plants with a total of 1,044 images, collected through direct photography and online sources. The data was split into training and testing sets with an 80:20 ratio, and offline data augmentation was applied to enhance data variability, including rotation, flipping, translation, and scaling. The training process was conducted without real-time augmentation since all augmentations had been performed offline. A baseline model (without augmentation) was first trained to obtain initial performance. Subsequently, ten experiments were conducted with different fine- tuning strategies, ranging from freezing all layers to unfreezing the last 70 layers. The training utilized callbacks such as EarlyStopping, ModelCheckpoint, and ReduceLROnPlateau to prevent overfitting and optimize performance. The evaluation results show that the baseline model without augmentation achieved an accuracy of 73.89%, while the best-performing model was obtained in the tenth experiment using a fine-tuning strategy that unfreezes the last 70 layers, adds two dense layers (256 and 128 neurons), and applies two dropout layers (0.5 each). This model achieved the highest accuracy of 84.07% on the test data. These findings demonstrate that data augmentation combined with structured fine-tuning significantly improves image classification performance for Acehnese herbal plants. Keywords: ResNet50V2, fine-tuning, data augmentation, Acehnese herbal plants, image classification