Emmanuel Damola Alofe: Geophysical and Machine Learning-based Enhancement for Characterisation of Urban Underground Space: Applications to Stockholm and Beyond

Abstract

As cities expand underground infrastructure to meet population growth and low-carbon goals, urban underground space (UUS) development remains largely reactive and poorly quantified. This thesis addresses three challenges: (1) the absence of a strategic planning framework for Stockholm’s underground; (2) a lack of scalable, geophysically grounded indicators for assessing UUS utilisation; and (3) limited resolution in airborne geophysical data of urban areas, constraining near-surface characterisation. Three studies employ geophysics, urban planning, and machine learning to support sustainable UUS development. Paper I establishes the planning context, revealing that Stockholm’s underground develops under a "first-come, first-served" principle. It also identifies fragmented geophysical data, inadequate 3D property rights, and poor cross-sectoral coordination as primary barriers. Paper II develops two geophysical proxy indicators for UUS use, especially when access to infrastructural information is unavailable or denied: analytic signal density and analytic signal per capita. Derived from 1995 airborne magnetic data targeting near-surface anthropogenic structures, both correlate with Stockholm’s population densities (r = 0.88 and r = −0.69), similarly to conventional indicators. A 3D magnetic susceptibility inversion model predicted the depths of known infrastructure to 20 m accuracy, while projections to 2023 estimate a 20 – 60% increase in UUS utilisation in central Stockholm, necessitating proactive planning. Paper III benchmarks three deep learning (DL) architectures against conventional downward continuation (DC) for high resolution and denoising of multi-altitude magnetic grids. Results show that DL models degrade gradually as acquisition height increases, while DC degrades rapidly—though DC excels when input data quality is already sufficiently good. The transformer-based DL variant most consistently matches or surpasses DC by capturing short- and long-range spatial dependencies. Together, these studies form a cohesive framework: identifying planning gaps, developing geophysical tools to quantify underground use, and exploring data-quality enhancements. This multidisciplinary framework is applicable to other cities where geophysical data of reasonable resolution are available.