Electronic Theses and Dissertation

Sejalan dengan agenda Sustainable Development Goals (SDGs) target 11.1 yang berfokus pada akses hunian yang layak, aman, dan terjangkau, diperlukan metode pemantauan kelayakan jenis material bangunan secara otomatis, efisien, akurat, dan berbiaya rendah. Penelitian ini mengangkat use case kecerdasan artifisial pada bidang perumahan dengan tujuan mengembangkan prototype sistem identifikasi otomatis material bangunan tempat tinggal berbasis data geospasial dengan memanfaatkan teknik deep learning. Dataset dikumpulkan dari sembilan kabupaten/kota di Indonesia menggunakan API Google Maps dan Google Street View. Setelah melalui tahap pembersihan diperoleh 1.475 gambar atap dan 1.218 gambar dinding yang kemudian dilakukan pelabelan dalam tujuh kelas atap yaitu beton, genteng, seng, asbes, kayu/sirap, ijuk, dan jerami/daun/rumbia serta lima kelas dinding yaitu tembok, plesteran kawat, kayu/papan, anyaman bambu, dan bambu. Kemudian diterapkan tujuh teknik augmentasi pada dataset pelatihan. Eksperimen dirancang berdasarkan jenis dataset (atap dan dinding), perlakuan augmentasi (original dan augmentation) ,dan arsitektur model deep learning (ResNet-50 dan MobileNetV3) dengan pendekatan transfer learning. Hasil penelitian menunjukkan bahwa performa terbaik diperoleh pada model ResNet50 yang dilatih dengan dataset augmentation pada masing-masing jenis dataset, dengan akurasi 97.67% pada dataset atap, dan 85.08% pada dataset dinding. Model yang telah dikembangkan kemudian diimplementasikan dalam antarmuka web berbasis Streamlit untuk mendukung pemantauan kelayakan bangunan secara interaktif.

In line with the Sustainable Development Goals (SDGs) 11.1, concerning access to adequate, safe, and affordable housing, there is a need for methods to monitor the adequacy of building materials in an automatic, efficient, accurate, and cost-effective way. This study implements an artificial intelligence use case in the housing sector by developing a prototype system for the automatic identification of residential building materials based-on geospatial data using deep learning techniques. The dataset was collected from nine regencies/cities in Indonesia using the Google Maps API and Google Street View API. After the data cleaning process, a total of 1.475 roof images and 1.218 wall images were obtained and labeled into seven roof material classes (concrete, tiles, zinc, asbestos, wood/shingle, palm fiber, and hay/leaves/rumbia) and five wall material classes (concrete, plastering of wire, wood/plank, woven bamboo, and bamboo). Seven data augmentation techniques were subsequently applied to the training dataset. Experimental scenarios were designed based on the dataset type (roof and wall), augmentation strategy (original and augmented), and deep learning architecture (ResNet-50 and MobileNetV3) using a transfer learning approach. The results indicate that the best performance was achieved by the ResNet50 model trained with augmented data for each dataset, achieving an accuracy of 97.67% on the roof dataset and 85.08% on the wall dataset. The developed model was then implemented in a Streamlit-based web interface to support interactive monitoring of building adequacy.