DEVELOPMENT OF BUILDING DETECTION SYSTEM FROM SATELLITE IMAGERY ON A WEB-BASED APPLICATION

• Main Author: Syafira, Beatrice
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/73904
• Institution: Institut Teknologi Bandung

Indonesia is a country that has high disaster potential. With a large population and rapid economic development, Indonesia has a high potential for losses if a disaster occurs. One of the potential losses caused by a disaster is damage to buildings and structures. Therefore, planning policies and developing financing plans is necessary to carry out disaster mitigation. This can be done by conducting disaster risk modelling as a disaster mitigation effort. A lot of supporting data is needed to carry out disaster risk modelling. One of which is the mapping of buildings in an area. The obstacle experienced so far is that manually mapping buildings in an area takes a lot of time and effort. For this reason, efforts are needed to automate building detection to solve this problem. AI technology is used in this final project to carry out automatic building detection from input in the form of satellite imagery. This final project will compare the Mask R-CNN model and the YOLOv5 model for building detection. This machine-learning model will detect buildings from input in satellite imagery. In addition, this machine learning model will obtain geolocation data and the address of each detected building. A comparison was made between the model obtained by the transfer learning process and the pretrained model. Both models will be tested by comparing the value of mAP as a parameter. The model with the highest mAP value will be taken as the model used for the building detection system. This building detection system will be implemented in a web- based application by applying a pre-selected model to carry out building detection. In addition, a post-processing optimization process was also carried out with the Non-Maximum Suppression algorithm in an effort to improve the detection quality of the previously selected model. The test results found that the pretrained Mask R-CNN model had the highest mAP score among the other models. Furthermore, geolocation data can also be obtained in the form of latitude and longitude as well as the address of the detected building. The results show that the optimization carried out with the NMS algorithm has yet to increase the quality of building detection. In addition, the web-based application can also receive input in the form of satellite imagery. This web-based application can also classify buildings by integrating with a building classification system. So, iv this application can provide the results of detected buildings in satellite imagery and the type of buildings detected.