Machine Learning Approaches for Building Inventory Characterisation

Kurzfassung

Accurate building inventories and relevant information are essential for sustainable urban governance, thereby contributing to achieve the Sustainable Development Goal 11, which focuses on sustainable cities and communities. Acquiring building information manually in complex built environment with densely populated buildings is impractical and may not guarantee the high thematic and spatial detail necessary for real world applications. Recent advancements in remote sensing technology, coupled with the availability of highresolution satellite and drone-based imagery through open access channels and as machine learning (ML) and deep learning (DL) continue to advance, showcasing their capability to recognise complex patterns, new opportunities for interpreting various surface features on the Earth are made possible. A growing body of literature emphasises the use of both traditional ML and DL for building footprint extraction, yet available and accessible literature reveal limited application of DL in urban building characterisation. This study involves exploring and implementing Random Forest (RF) as machine learning model and dense neural network (DNN) as deep learning model in the context of multi-class building characterisation encompassing six classes. A total of 35 geometric and distribution features calculated using VHR imagery and OpenStreetMap data are used to train the model. The experiments show that overall accuracy of RF (79.9%) is higher than that of the DNN (71.9%). Upon closer examination and comparison of diagonal elements, representing the number of correctly classified samples for each class, it is found that the DNN outperforms RF in correctly classifying more instances for four classes. Further the recall rate using DNN is greater for four classes- ‘Building block in closed construction’, ‘Detached building block’, ‘Free standing individual building’, and ‘Garage’, in comparison to that of RF. Implementation of the DNN and comparison with traditional machine learning algorithm- RF provide additional scientific contribution, especially in situation where there is limited use of deep learning algorithms in the building characterisation.