The Application of Deep Learning Models for the Extraction of a Historical Street Map of Sydney

Abstract

This thesis investigates the application of deep learning models for the extraction of historical street maps of Sydney, focusing on two primary methodologies: the proposed Edge Extraction method with a novel hyperparameter optimisation approach and the existing GIS Labelling Overlay (GLO) method (Huang, 2023). The research aims to improve the accuracy and efficiency of road network extraction from historical maps, a task traditionally accomplished through labour-intensive manual methods. By leveraging Convolutional Neural Networks, specifically the U-Net model and the DiResNet ensemble, this study evaluates the performance of these models in terms of accuracy, F1 Score, and the Intersection over Union metric. Additionally, this thesis introduces an innovative approach to hyperparameter optimisation. It involved reducing the hyperparameter permutation space with Random Search to run a complete search on the optimally reduced space with the Grid Search algorithm. This was applied to the Edge Extraction technique in Experiment 1, contributing to significantly better performance than the GLO method.

The experiments reveal a clear superiority of the proposed Edge Extraction method over the GLO method, with significant improvements of the F1 Score and IoU, thereby demonstrating its robustness and reliability. Further, both the U-Net and DiResNet models exhibit high performance in road network extraction, with U-Net achieving slightly higher accuracy and F1 Score while DiResNet shows better Intersection over Union metric. These results indicate that while both models are effective, each has its strengths in different aspects of performance. Ultimately, U-Net had a clear edge in training time which is practically advantageous with regards to model optimisation and testing.

Future research should focus on refining the Edge Extraction method to automate the removal of non-road, extraneous objects from the extracted images, aiming for a fully automated data labelling process. This research contributes to the field of digital cartography and historical analysis, offering enhanced tools for studying urban development and human settlement patterns through improved road network extraction techniques.