An improved segmentation and classification method for building extraction from RGB images using GEOBIA framework
Loading...
Authors
Hossain, Mohammad D.
Date
Type
thesis
Language
eng
Keyword
GEOBIA , Building Extraction , UAV images , Image Segmentation , Shallow Classifiers , Deep Tabular Classifiers
Alternative Title
Abstract
Cities worldwide face a range of interconnected challenges, including environmental issues, housing shortages, transportation problems, and crime rates. These challenges are closely tied to land use and land cover, with buildings playing a crucial role. A systematic and repeatable method for quantifying buildings is essential to tackle these issues effectively. Remote sensing offers the advantage of providing high-resolution data over large areas, making it a valuable tool. In remote sensing applications, the near-infrared (NIR) band, together with the visible spectrum (RGB), provides abundant information about ground objects. However, low-cost remote sensors like UAVs and aerial photographs encounter difficulties in accurately differentiating buildings from other features due to the lack of an NIR band. Geographic Object-Based Image Analysis (GEOBIA) has been developed for analyzing high-resolution remote sensing images, but it faces challenges in segmenting and classifying buildings. This research addresses three significant challenges in building mapping. Firstly, it explores a segmentation method designed to produce optimal segments for buildings. The study implemented a hybrid segmentation approach combining edge- and region-based methods, providing better accuracy. This hybrid method ensures intra-segment homogeneity within objects and inter-segment heterogeneity between objects, surpassing existing published methods. Following segmentation, classifying segments using machine learning (ML) classifiers is a crucial step in GEOBIA. Shallow learning (SL) classifiers are commonly used, but the effectiveness of Deep Learning (DL) classifiers in GEOBIA classification has not been thoroughly tested. This research simultaneously implemented SL and DL classifiers, revealing that DL classifiers like Gated Additive Tree Ensemble (GATE) and Neural Oblivious Decision Ensemble (NODE) perform better in many cases. However, these classifiers still struggle to differentiate buildings from other objects, such as parking lots, driveways, and roads. Lastly, a target-based building extraction method was implemented, focusing on buildings of different colors. This method demonstrated improved accuracy compared to alternative approaches. The significance of this research lies in addressing the fundamental issues within the contemporary GEOBIA method. The methods and results presented in this study can be readily applied to other geographic areas and image settings, offering practical solutions beyond specific regions. By following the studies outlined in this research, city authorities worldwide can enhance their efficiency in addressing urban challenges.