Novel machine learning techniques to generate high spatio-temporal resolution urban climate observations for climate adaptation

As cities worldwide expand, urban areas are experiencing rising air and surface temperatures. These temperature increases exacerbate the impacts of climate change, threatening the health, well-being, and livelihoods of urban populations. To mitigate these risks, urban planners and policymakers must adopt climate adaptation and resilience strategies that account for the

specific conditions of each city. This requires access to accurate, high spatio-temporal resolution urban climate data. This data can help assess risks, hazards, exposure, and vulnerabilities to climate impacts. However, many cities, particularly in the Global South, lack the necessary resources or infrastructure to gather this data in sufficient detail, limiting their

ability to plan for climate change effectively. This research proposes an innovative approach to address this challenge by integrating crowdsourced urban climate observations from Citizen Weather Stations (CWS) with satellite and remote sensing data. The aim is to generate high-resolution urban atmospheric dynamics

(500m) observations in data-scarce cities. The research methodology will focus on three cities located in different climate zones, allowing for a comparative analysis of urban climates across diverse environments. The research will develop a full statistical analysis of existing data sources and employ machine

learning techniques, particularly deep learning models, to develop high-resolution datasets generated by the integration of CWS and remote sensing data. Finally, novel transfer learning techniques will be used to apply models trained on cities with rich datasets to those with sparse or incomplete data. This approach would allow for more accurate predictions of urban

climate behaviour globally, even in data-poor areas. A key novelty of this thesis lies in the use and integration of novel and rich urban datasets with advanced machine learning techniques. The project will use CWS data, which is crowdsourced and often hyperlocal, providing granular temperature, humidity, and other

meteorological observations. When combined with the broader-scale observations from satellites and remote sensing technologies, the result is a uniquely detailed view of urban climate dynamics, including hourly, daily and seasonal variability, heat waves, and other phenomena. Moreover, the development of the machine learning-based tool will offer cities

lacking comprehensive climate data a way to access detailed climate insights. This tool will be designed to integrate with existing urban planning frameworks, enabling cities to assess their climate risks, plan adaptive measures, and move towards sustainable, low-carbon urban environments. The findings will provide a foundation for cities globally to incorporate high-resolution

climate data into their planning and policy efforts, promoting more resilient, climate-adaptive urban growth. The tool developed from this research will be globally applicable to help cities, particularly those with limited data resources, to better understand their unique climate challenges and take proactive steps towards climate adaptation. This work aligns with the

EPSRC themes of manufacturing, circular economy, energy, decarbonisation, and resilience, offering a pathway to more sustainable, climate-sensitive urban planning.