Transforming urban planning through machine learning: A study on planning application classification using natural language processing

Planning for sustainable urban growth is a pressing challenge facing many cities. Investigating proposed changes to the built environment can provide planners and policymakers information to understand future urban development trends and related infrastructure requirements. It is in this context we have developed a novel urban analytics approach that utilises planning applications (PAs) data and Natural Language Processing (NLP) techniques to forecast the housing supply pipeline in Australia. Firstly, we implement a data processing pipeline which collects, geocodes, and filters PA data from local government websites and state planning portals to provide the first nationally available daily dataset of PAs that are currently under consideration. Secondly, we classify the collected PAs into four distinct urban development categories, selected based on the infrastructure planning and provisioning they require. We tested five model architectures and found that the fine-tuned DeBERTA-v3 model achieves the best performance with both accuracy and F1-score of 0.944. This demonstrates the suitability of fine-tuned Pre-trained Language Models (PLMs) in classifying PAs based on text in them. Finally, this model is applied to classify and map urban development trends in Australia’s two largest cities, Sydney and Melbourne across two time periods – between 2021 and 2022, and between 2023 and 2024. This mapping affirms a face-validation test of the classification model and demonstrates the utility of insights the generate for planning use case. Holistically, the paper demonstrates the potential for NLP to enrich urban analytics through the integration of previously inaccessible text data in planning applications into planning analysis and decisions.