Electronic Theses and Dissertation

Klasifikasi tipe rumah di Indonesia masih dilakukan secara manual sehingga bersifat subjektif dan kurang konsisten, yang dapat memengaruhi penerapannya dalam penentuan Uang Kuliah Tunggal (UKT), beasiswa, maupun bantuan sosial. Penelitian klasifikasi citra rumah yang ada umumnya berfokus pada gaya arsitektur atau menggunakan dataset luar negeri sehingga belum merepresentasikan karakteristik visual rumah di Indonesia. Penelitian ini bertujuan mengembangkan model klasifikasi tipe rumah Indonesia berdasarkan kategori sosial-ekonomi, yaitu rumah sederhana, menengah, dan mewah, melalui integrasi augmentasi data dan transfer learning menggunakan UTU-House Dataset. Dataset yang dikembangkan terdiri atas 3.000 citra tampak depan rumah yang telah dianotasi dan berasal dari 31 provinsi di Indonesia. Penelitian membandingkan lima arsitektur Convolutional Neural Network (CNN) pralatih, yaitu VGG16, ResNet50, DenseNet121, MobileNetV2, dan EfficientNetB0, melalui tiga skenario eksperimen, yaitu pelatihan dasar (baseline), augmentasi data, dan transfer learning dengan fine-tuning. Evaluasi model dilakukan menggunakan metrik accuracy, precision, recall, F1-score, confusion matrix, serta validasi 5-fold cross-validation. Hasil penelitian menunjukkan bahwa EfficientNetB0 memberikan performa terbaik, dengan accuracy dan macro F1-score meningkat dari 0.85 pada tahap baseline menjadi 0.92 setelah proses fine-tuning. Validasi menggunakan 5-fold cross-validation menghasilkan rata-rata validation accuracy sebesar 97.13% dengan standar deviasi 0.19%, yang menunjukkan kemampuan generalisasi model yang baik dan stabil. Sebuah prototipe berbasis Streamlit juga dikembangkan untuk mendemonstrasikan klasifikasi secara real-time. Penelitian ini memberikan kontribusi berupa dataset lokal baru, pendekatan klasifikasi yang andal, serta alat implementasi praktis untuk keperluan verifikasi visual.

House type classification in Indonesia is still conducted manually, making it subjective and inconsistent, which can affect its application in determining single tuition fees (UKT), scholarships, and social assistance. Existing image-based house classification studies generally focus on architectural styles or rely on foreign datasets, and therefore do not represent the visual characteristics of Indonesian houses. This study aims to develop a classification model for Indonesian house types based on socioeconomic categories, namely simple, middle, and luxury houses, through the integration of data augmentation and transfer learning using the UTU-House Dataset. The dataset consists of 3,000 annotated front-facade house images collected from 31 provinces in Indonesia. Five pretrained Convolutional Neural Network (CNN) architectures, namely VGG16, ResNet50, DenseNet121, MobileNetV2, and EfficientNetB0, were compared through three experimental scenarios: baseline training, data augmentation, and transfer learning with fine-tuning. Model evaluation was carried out using accuracy, precision, recall, F1-score, confusion matrix, and 5-fold cross-validation. The results show that EfficientNetB0 achieved the best performance, with accuracy and macro F1-score improving from 0.85 in the baseline scenario to 0.92 after fine-tuning. Validation using 5-fold cross-validation produced an average validation accuracy of 97.13% with a standard deviation of 0.19%, indicating good and stable model generalization. A Streamlit-based prototype was also developed to demonstrate real-time classification. This study contributes a novel local dataset, a reliable classification approach, and a practical implementation tool for visual verification purposes.