DEEP LEARNING UTILIZATION FOR BUILDINGS AND ROAD NETWORKS EXTRACTION FROM ORTHOPHOTO (STUDY CASE: SUMEDANG REGENCY)

• Main Author: Faisal Anshory, Muhammad
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/69043
• Institution: Institut Teknologi Bandung

Object extraction from remote sensing imagery is commonly still done manually through visual interpretation by the operator. It can take a long time and is expensive as the scale of the mapping increases. Deep learning technology has unlocked the potential of automatic object extraction from remote sensing imagery products such as orthophotos. However, deep learning is prone to bias that results in the underrepresentation of a particular group. In accordance with the existing problems and potentials, this study aims to develop a deep learning model for the extraction of buildings and road networks and identify the bias in the model in two study areas in Sumedang Regency. In Mask R-CNN-based building extraction model, identification of bias is carried out on the categories of regular and irregular settlements. Meanwhile, in U-NET-based road network extraction model, identification of bias is carried out on road category 1 which includes ‘arteri’ also ‘collector’ roads and class 2 which includes ‘lokal’ also ‘lingkungan’ roads. The metric evaluation of the building model in the two areas shows that regular settlements give higher F1 scores of 0.77 and 0.69 compared to irregular settlements which are only 0.72 and 0.66. On the other hand, the road network model evaluation metric shows that class 1 roads give better intersection of union values of 0.76 and 0.84 compared to class 2 which are 0.74 in both areas. It shows that deep learning can be used to extract buildings and road networks, but manual correction is still required. In addition, the evaluation metric of the road network extraction model in different categories shows more consistent results than the building extraction model.