DEVELOPMENT OF BUILDING CLASSIFICATION SUBSYSTEM AND FRONTEND SUBSYSTEM ON A WEB- BASED BUILDING DETECTION APPLICATION

• Main Author: Zahra Layungsari K, Roro
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/74072
• Institution: Institut Teknologi Bandung

Indonesia's location in a cluster of volcanoes makes its territory vulnerable to natural disasters caused by tectonic plate collisions, such as earthquakes. One of the objects affected when the earthquake hit Indonesia was buildings. Damages caused to buildings can be mitigated by disaster risk modeling that can compare conditions before and after a disaster. This comparison can provide information related to the scale of the disaster by looking at the differences in the condition of the buildings before and after the disaster. This knowledge can then be used by the authorities to estimate the number and type of buildings affected by a disaster, as well as the required reconstruction effort. One of the processes needed in disaster risk modeling is building mapping, which includes detecting and classifying buildings. However, this process takes a long time, especially for large areas. Therefore, we need a scheme that can automate this process. This research develops a building detection and classifier system integrated into a web-based application. This system receives input from satellite imagery and processes it into an output containing the results of building detection. The results of this building detection are then integrated with the building classification subsystem. The building classification subsystem assigns buildings based on their occupancy type, residential and non-residential. This subsystem utilizes the best- performance machine learning model, which is selected based on the results of comparing the EfficientNetB7, MobileNetV2, and VGG16 models. These models were trained using the Beauty and UK Building Facade datasets and optimized using the Adam optimizer. The training processes performed on each model produce precision, recall, and F1 scores which are used as review materials in comparing the performances of the three models. The results obtained from this test state that the MobileNetV2 model trained by the Beauty dataset is the model with the best performance among other models and dataset variations. The web-based application development stage carried out in this study includes the process of forming the system's user interface. The user interface is designed by creating frontend components that define the structure and design of the web pages. This component is tested to test its functionality and responsiveness to changes and input from users. The results of this test indicate that the frontend component can run properly.