DEVELOPMENT OF MACHINE LEARNING APPLICATION FOR AUTOMATED BUILDING DETECTION AND CLASSIFICATION FROM SATELLITE IMAGERY AND STREET VIEW IMAGES

• Main Author: Pambudi Adhiluhung, Gelar
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/70445
• Institution: Institut Teknologi Bandung

The rapid growth of infrastructure and population in Indonesia causes increases losses due to disasters every year. One of the objects affected by a disaster is the buildings assets in a certain area. A good disaster risk model is needed to plan and finance disaster mitigation. One of the challenges faced in the disaster risk modeling process is that the manual mapping of buildings in an area often takes a long time, especially for large areas. Building classification AI applications were developed to simplify and speed up the process. Users only need to enter satellite imagery as input to the application and will get results in the form of building data extracted from the satellite image of a certain area The applications are developed in the form of web applications to provide easy access for users. The application utilizes two types of machine learning model, namely model to detect buildings on satellite images and model to classify buildings based on street view images of buildings. The building detection model uses Mask R-CNN based model, while the building classification uses a CNN-based model. The building detection model will detect the existing buildings in the satellite image and obtain the geolocation coordinates of the building. The geolocation data of the building is then used to obtain a street view image of the building which will be used for the building classification process. The development of the machine learning model is carried out using the transfer learning method using several different base models. The transfer learning result of building detection model is then tested by comparing the mAP of each model using different satellite images. The transfer learning result of building classification model was carried out by comparing the precision, recall, and f1-score of each building category in each model. Application testing is carried out to test the functionality of each application component, namely the frontend and backend. The test results showed that the application was successfully created by integrating machine learning model for building detection and building classification. The application can receive input in the form of satellite images and carry out the detection and classification process automatically. The application can provide results in the form of images of building mapping on satellite imagery and data on types of buildings detected and their locations.