Foundation model technology for Urban Heat Islands

Abstract

Assessing and understanding the urban scale impacts of extreme climate events is a global necessity. Risks associated with heat, where intra-urban dynamics and rural/urban boundary conditions greatly impact its distribution, are of particular interest as the evolution of climate change and urbanization persists. Characterizing Urban Heat Island (UHI) effects is dependent on the availability of high-resolution near-surface air and land surface temperature datasets. Along with the rapid rate of urbanization across African cities, significant levels of informality bring about escalating adverse health consequences, which are further exacerbated by UHIs and extreme heat events. Highlighting the insufficient understanding between heat related health outcomes, exposure characterisation, vulnerability of populations, and potential solutions within African urban contexts.  UHIs are defined by the higher temperatures experienced in built urban environments as opposed to their surrounding rural landscapes. This difference results from the varying absorption and reflectivity capacities of urban materials, reduced natural cover (such as water bodies and vegetation), increased anthropogenic activities, and urban geometry, amongst others. Intra-urban distributions in temperature are also significantly impacted by these factors. With world cities becoming more populous and increasing in size, the need for high-resolution urban temperature is essential.  The increased heat in the urban areas has several negative impacts on city operations, resource management, health and livelihoods. For example, intense UHI effects in dense cities can increase energy consumption as a result of increased air-conditioning and cooling needs. Another example reported excess costs for AC repair and operation due to UHI effects approaches $478 million annually for Phoenix (USA). In addition, the higher-than-average temperature has a large impact on human health; especially affecting infants, individuals with co-morbidities, the elderly and other vulnerable populations. The increased impact of UHI effects on human health can lead to increased strain on health care systems. These effects are significantly amplified during heat-waves which intensify the UHI effects. A study in the West Midlands (UK), attributes 50% of the heat related deaths in the 2003 summer heat waves to UHI’s. One challenge in addressing the impacts of UHI is the lack of timely high-resolution, urban scale temperature data. Such data would allow for improved urban design and planning, as well as the development of early warning systems and disaster response mechanisms.  Existing urban temperature modelling approaches are largely dependent on heterogeneous sets of data, expert user input, high computational resource requirements, significant parameterization, and are biased towards data rich regions. This increases the barrier to wide-scale adoption by urban planners, disaster relief organizations, and other relevant stakeholders. AI approaches as an alternative, have been shown to require less a priori information, less computational resources, do not necessarily need heterogeneous sets of data, and do not require parameterizations. Geospatial Foundation Models advance the development of current AI models by training the models on large corpuses of unstructured geospatial data. The advantage being that these models are highly generalizable to various downstream tasks, such as urban scale temperature mapping.  We present our findings from the application of the IBM Earth Observation Foundation Model, “Prithvi”, to fine-tune a global, high-resolution, land surface temperature (LST) model, reporting a mean absolute error measure less than 1.7 °C. The model was fine-tuned on a combination of remote sensing (HLS Landsat30) and reanalysis (ERA5 Land) climate datasets for global cities of varying climate zones over the period 2013 - 2023. The developed model incorporates a SWIN transformer architecture which offers a hierarchical structure to account for long range dependencies. Performance was benchmarked against a U-Net architecture for the regression task of land surface temperature prediction. Our results indicate high correlation between predicted and measured values of LST, with enhanced capabilities to inference on unseen cities of interest. The model is able to accurately capture temperature variations across structural features within a city’s urban landscape. To iterate on this study, we will consider auxiliary climate variables such as; Normalised Difference vegetation Index, Urban Index. Bare Soil Index, Automated Water Extraction Index, digital elevation as well as reanalysis solar radiation components.  The model developed is beneficial for UHI detection with the ability to contribute to urban scale heat hazard mapping and forecasting. Further application can be extended to; integration with health outcomes datasets for adverse heat related health associations modelling, impact assessment on built environment, specifically critical infrastructure, urban scale city planning and development, to name a few. Such applications are crucial for the development of an Early Warning Systems across African cities. An existing partnership between IBM Research Africa and the Heat and Health African Transdisciplinary Center (HEAT Center) aims to develop innovative solutions for mitigating the health impacts of climate change in Sub-Saharan Africa. The impact of our work extends beyond the development an Early Warning System within the HEAT Center partnership as we are able to present global context on the detection and characterisation of UHIs.