Asbestos roofing AI change detection
Jul 2024
Accurate detection of changes in asbestos roofing is vital due to help manage the significant health risks associated with asbestos exposure.
Jul 2024
)
Redcliffe, QLD AU
In his role as Infrastructure Planner at the City of Gold Coast, Matt Abbasi is part of a wider professional team that uses Nearmap high-resolution aerial imagery to craft spatial analytics-driven solutions that benefit communities and the environment, such as visually inspecting property parcels, planning, and monitoring projects and assets.
As an academic, Matt is also completing a PhD focusing on utilising innovative technologies for automation in environmental management.
In collaboration with various institutions including the Griffith University Cities Research Institute, and the University of New South Wales School of Civil and Environmental Engineering, Matt and his research group created a hybrid object-based deep learning framework with post-classification structure: an AI-based model that is capable of learning both spatial and temporal features of aerial images for multi-temporal change detection of asbestos roofing.
“As someone who has worked with different aerial images, the temporal resolution of Nearmap aerial imagery is great because the archive goes back to 2009 and it’s consistent across many urban areas, which is really important.”
Matt Abbasi, PhD candidate, Griffith UniversityAs input data, the study used high-resolution Nearmap aerial imagery across seven time points: 2010, 2012, 2014, 2016, 2018, 2020, and 2022 in five metropolitan areas in Australia.
The image below shows (a) the locations of study areas, (b) the extent of Southport-North, QLD (c) the extent of Braybrook, VIC (d) the extent of Redcliffe, QLD (e) the extent of Balga-Mirrabooka, WA (f) the extent of Blacktown-South, NSW.
)
These five residential built-up areas within the Australian Bureau of Statistics (ABS) unit of analysis, Statistical Areas Level 2 (SA2) – medium-sized general purpose areas representing communities that interact together socially, and economically.
This created a unique imagery dataset so the model could be trained and tested across a diverse spectrum of residential roofing, encompassing various building sizes, shapes, and roofing materials. Key to this was the accuracy of the Nearmap aerial imagery of the roofs across the seven points of time between 2010 and 2022.
Hybrid deep-learning models integrate well-documented, high-performance classifiers, eliminating the need for additional image preprocessing or complex deep learning models to address challenges with tracking changes over time.
The study group worked with a Nearmap high-resolution vertical (top-down 2D) image of each roof across each of the seven time points, and built a patch sequence for each roof with an index, which was used as the input for the deep learning models using feature extractors to identify spatial and temporal features.
Asbestos roofing often has spectral characters similar to other materials, such as certain types of concrete roofing tiles or asphalt pathways, making detection challenging.
The image below is a depiction of a roof with Super Six asbestos cement sheeting across seven time points, illustrating the geometric and illumination inconsistencies, as well as the impact of overgrown trees and shadows. The red polygon marks the building’s location in 2010, emphasising the misalignment of the building positions in subsequent years.
)
Starting with that ground truth – the labelled data – the team trained the model to look at asbestos roofing across time points to create multi-temporal change maps, showing asbestos roofs at the outset 2010 and identified changes across each of the seven time points. ‘Change’ was defined as development on the roof, recognising a change of roofing material.
The study framework handles inconsistencies such as illumination variation in temporal data without preprocessing, and used legislative information to enhance the classification accuracy.
The hybrid deep learning models achieved an overall accuracy of 96%. In October 2023, the Geospatial Council of Australia awarded Matt Abbasi a Geospatial Excellence Postgraduate Student Award.
The image below shows the predicted asbestos roofs (the red polygon shows predicted asbestos roofs in 2010); (b) illustrates the creation of the change map based on label variations in roofs over temporal years; red polygon represents asbestos roofs in 2010, indicating no change where repeated in the next years, green represents asbestos roofs changed in 2014, yellow represents asbestos roofs changed in 2016, orange shows asbestos roofs changed in 2020, and blue shows asbestos roofs changed in 2022.
)
Based on the multitemporal study on roof developments, the study group will launch its next stage of research: to train a model that will forecast when existing asbestos roofs are most likely to be replaced.
To find areas more probable to have asbestos roofing, the study followed a multi-criteria analysis to identify areas that were more probable to have significant levels of asbestos roofing.
The models can be used to detect asbestos roofs in any areas across 75% of New Zealand’s population covered by Nearmap aerial imagery.
“So when the resolution is high enough so we can detect those objects much better than satellite images. And that’s the reason why, for many urban studies like change detection of buildings, aerial images are the preferred raw data that we can use rather than satellite images,” said Matt Abbasi.
The image below shows: (a) labelled buildings with blue, green, and red polygons representing asbestos roofing, non-asbestos roofing, and non-labelled buildings, respectively, (b) layout 1 in 2010, (c) layout 1 in 2014, (d) layout 1 in 2018, (e) layout 1 in 2022.
)
This study is a reproducible framework for any application that needs a temporal change map of specific objects at the building level, including councils, industry partners, health agencies and more.
As new deep learning models are created to analyse complex data from a spatial and temporal perspective, communities will benefit from more accurate insight-driven solutions. The outcome of this research will provide insights for asbestos management strategies.
)
Read more about Multi-temporal Change Detection of Asbestos Roofing: A Hybrid Object-Based Deep Learning Framework with Post-Classification Structure, published in the Journal of Remote Sensing Applications: Society and Environment.