IBM EOFM Enabling Seamless Characterization of Surface Urban Heat Islands (SUHIs)

Abstract

Quantifying the impacts of extreme climate events is a global necessity. Risks associated with heat and how they vary within urban areas and between urban and rural areas, are of particular interest as the pace of climate change continues to quicken. African cities with rapid rates of urbanisation, leading to significant levels of informality, are a concern in terms of escalating adverse health consequences associated with extreme heat events. One approach to understanding the heat hazards and their variation along the urban-rural gradient is to characterize the Urban Heat Island (UHI) effect. Physics-based modelling of the UHI depends on the availability of high-resolution air, land surface temperature and other datasets required for energy balance and heat flux modelling, along with high performance computing and expertise required to develop these high fidelity models. The inequities in the availability of high resolution climate data and limited scalability of physics-based UHI models is thus a challenge. We consider Artificial Intelligence (AI) approaches, namely Geospatial Foundation Models (GFM) as an alternative for developing a scalable UHI model that leverages existing disparate climate and multispectral satellite data. GFMs are trained on large corpuses of geospatial data, making them highly generalizable to various downstream tasks, such as urban scale temperature mapping. We leveraged the IBM Earth Observation Foundation Model, “Prithvi”, a GFM, fine-tuning this model on global, high-resolution remote sensing data (HLS Landsat 30) and reanalysis (ERA5 Land) climate datasets for cities of varying climate zones over the period 2013 – 2023, to predict land surface temperatures (LST). The fine-tuned model incorporates a SWIN transformer architecture. The performance of the model in predicting LST was assessed against a U-Net model. Our results indicate high correlation between predicted and measured values of LST, with mean absolute error less than 1.7°C, and enhanced capabilities for inference on unseen cities. Our model is able to accurately capture temperature variations across different land cover types within urban areas. The model developed is beneficial for UHI detection with the ability to contribute to urban scale heat hazard mapping and forecasting. Further, the outputs of the model can be further pot-processed and linked with health outcomes datasets for quantifying the impact of heat on health and the development of Early Warning Systems. Other applications include assessing the impact of heat on the built environment, specifically risks to critical infrastructure, and inform urban planning decisions related in the context of adapting cities to mitigate against the effects of UHIs.