Data

[D5-02] Advancing Building Extraction in Thailand using the YOLOv8 Segment Model on Open-source Data across Diverse Land Use Types

*Bhanu Prasad CHINTAKINDI¹, Yoshiki OGAWA¹, Yoshihide SEKIMOTO¹ (1. The University of Tokyo)

Keywords:YOLOv8 Segment Model, Global Building Extraction, Land-Use Types, Mask R-CNN Base Model, Open-source Data

The existing global building extraction models perform well in only a few land-use types. Within each land-use type, it has limitations while serving in regions like low-rise buildings, high-rise buildings, buildings surrounded by vegetation, and dense neighborhoods. This research addresses these challenges by leveraging the versatile YOLOv8 segment model, renowned for its speed, accuracy, and user-friendly design. The methodology involves fine-tuning the YOLOv8 segment model to develop individual land-use-based building models for Commercial, Urban, Suburban, Built-Up, Rural, and All Land Use Types Combined categories in Bangkok province, Thailand. Performance and generalizability are assessed by evaluating the models on unseen regions encompassing various land-use types and cross-validating with the Mask R-CNN base model trained in Japan. The findings advance building extraction techniques and provide insights into Thailand's land-use challenges.