Influence of image source type and spatial resolution on deep learning-based automated green roof recognition

Climate change is intensifying the frequency and severity of extreme hydroclimatic events such as heat waves, droughts, and heavy rainfall. These effects are particularly pronounced in urbanized areas with extensive paved surfaces and limited vegetation. Rising temperatures exacerbate urban heat islands (UHI), while heavy rainfall can overwhelm drainage systems, increasing the risk of flooding and combined system overflow. As one of the most widely applied blue-green infrastructure in urban regions, green roofs offer a promising solution to these challenges. By mitigating UHI effects and enhancing stormwater management, they can significantly contribute to urban climate resilience. Green roofs can reduce summer heat gain in buildings by up to 31% and retain an average of 87% of rainfall, with a substantial portion returned to the atmosphere. Despite their potential, comprehensive assessments of green roof adoption and effectiveness remain limited, partly due to a lack of accessible, comprehensive data on their prevalence and performance. Additionally, data tracking the development of green infrastructure over time is scarce, hindering the evaluation of policies and incentives aimed at promoting their implementation.

To address this gap, previous work by Wu and Biljecki developed “Roofpedia”, an open-source deep learning algorithm for green roof mapping and urban sustainability evaluation using satellite imagery. This model employs a convolutional neural network (U-Net) for image segmentation and has been successfully applied to satellite imagery and aerial orthoimagery data from different cities worldwide. Satellite imagery and aerial imagery collected with ad hoc campaigns can, however, be characterized by very different spatial resolution.

Acknowledging that different types of images and image resolutions can affect the feasibility and accuracy of automated green roof recognition, this research retrains and evaluates Roofpedia using imagery data of Berlin (Germany), investigating quantitatively how image platform type and spatial resolution affect the accuracy of automated green roof detection accuracy. Preliminary results show that green roof classification accuracy degrades substantially when the algorithm trained on orthoimagery with a 0.2 m/pixel resolution is transferred for application onto satellite imagery with a spatial resolution of 3 m/pixel, hampering the prediction of green roofs at this resolution. Further research will investigate how green roof classification capabilities degrade for intermediate resolutions, possibly identifying a feasibility range, along with different algorithm training and testing strategies considering combinations of image sources. This research ultimately aims to enhance the effectiveness of automated tools for green roof mapping, providing actionable insights to support urban planning, policymaking, and the broader adoption and monitoring of green infrastructure as a climate adaptation strategy.