Towards Bridging GIS and 3D Modeling: A Framework for Learning Coordinate Conversion Using Machine Learning

Simulating what-if scenarios in 3D environments has become popular due to its potential to improve planning and operations for large-scale, complex events, such as the annual Hajj. Despite the extensive availability of spatial data through Geographic Information Systems (GIS)—information systems specialized in collecting, organizing, and analyzing geospatial information—these resources remain underutilized for such use cases. Three-dimensional modelers create simulations for these use cases and build them from scratch despite the availability of GIS data, which can be tedious and error-prone. We raise the question of whether it is reasonable to use machine learning (ML) algorithms to learn coordinate conversion systems, laying the foundations for accelerating the construction of 3D environments and achieving more accurate results. In this paper, we introduce a simple yet novel framework that facilitates learning coordinate conversion systems. Using our framework, we evaluate 35 ML models and provide a detailed analysis of their prediction performance, overfitting characteristics, and trade-offs in terms of time and model size. Notably, 14 of the ML models demonstrate near-perfect learning (i.e., R-squared very close to 1.0). Furthermore, we use grid search to find better parameter settings for underperforming models. Our results demonstrate that ML can help 3D modelers automate manual tasks and improve the efficiency of 3D modeling from GIS data.