Automated property parcel classification using very high-resolution land cover imagery and multi-modal deep learning: Effective city-scale monitoring for urban expansion outcome modelling

This study addresses the challenges of monitoring urban development where sprawl, infill and soft densification are prevalent. To accurately identify the spatial distribution and characteristics of different development types, a multi-modal deep learning model was developed to predict parcel-level development types using very high-resolution land cover maps and tabular data representing parcel attributes of surface cover, built form and dwelling counts. This model was compared to a range of supervised and unsupervised classification models and evaluated in two contexts: (i) within Perth, where training data was abundant, and (ii) when generalising models to other cities across Australia. To undertake this analysis, a ground truth dataset of almost 70,000 labelled parcels were sampled from Perth, Western Australia, with fine-tuning and test datasets produced for other Australian cities (Melbourne, Sydney, Brisbane and Adelaide). The multi-modal deep learning model achieved an overall accuracy of 94% when evaluated in Perth and overall accuracy ranging from 91% to 94% when evaluated in other cities. The model was more accurate in classifying visually distinct development types that are relevant to monitoring infill and densification and performed well in transferring to other cities. These results indicate that deep learning models and transfer learning workflows can be leveraged to develop generalisable urban development type mapping models. This approach opens opportunities for multi-city analyses to further understand the spatial and temporal dynamics of urban growth and densification, and provides data to guide and evaluate the effects of policy that seeks to balance the need for housing our growing populations with maintaining healthy natural environments.