Accuracy of Google’s Open Buildings Dataset in Characterising Diverse Building Configurations in South Africa

Accurate building footprint datasets are critical for urban planning, disaster management, and infrastructure development. Conventional methods for generating these datasets are labour-intensive, costly, and susceptible to human error. While AI-driven approaches offer scalable, cost-effective, and spatially explicit alternatives, the accuracy and reliability of global AI-derived datasets, such as Google’s Open Buildings (GOB), remain understudied, limiting their adoption in real-world applications. This study evaluates the performance of GOB—a deep learning-generated global building footprint dataset—in characterising complex building configurations across South Africa. A reference dataset was created through the manual digitisation of buildings from 0.3 m spatial resolution ESRI Wayback imagery across five study sites, selected to represent the country’s diverse building configurations: informal settlements (Alexandra), low-cost housing (Cosmo City), industrial areas (Kempton Park), a mining town with mixed land use (Rustenburg), and a rural area (Nongoma). A comparative analysis between GOB and the reference dataset was conducted using various accuracy metrics, i.e., Precision, Recall, and Intersection over Union (IoU). The results indicate that GOB achieved a high Recall of 90%, demonstrating robust building location detection capabilities. However, Precision was consistently low (≤50%), reflecting the frequent misclassification of other features as buildings. The IoU values ranged from 10% to 45%, indicating considerable variation in agreement between detected and reference buildings across the study areas, with poorer performance for irregularly shaped structures (e.g., informal settlements) compared to regular geometries. The dataset can be effective for applications prioritising building counts or density, such as flood risk assessments. However, it is inadequate for tasks requiring precise boundary delineation, such as rooftop runoff and photovoltaic solar potential estimation. This study provides the first systematic evaluation of the GOB dataset across diverse South African building configurations, highlighting its potential for large-scale urban analytics while detailing its limitations in geometric precision, thereby advancing the understanding and cautious application of AI-driven geospatial solutions in urban planning frameworks. Future efforts should prioritise refining AI models through enriched training data, particularly for complex building morphologies.