Urban Heat Island Estimation with Machine Learning

Abstract

Introduction:  Downscaling in earth or climate science involves transforming a low-resolution image or layer of a variable into a corresponding high spatial resolution. Gridded data obtained from global or sometimes regional numerical weather or climate models, are fairly coarse and often unable to resolve the spatial heat signatures and variability that are important for characterizing the Urban Heat Island (UHI) effect.  Specialized physics-based urban models exist that are capable of simulating urban-scale temperatures for any city, but these either require substantial computational resources and time or expert user inputs, or they are often not open access, which limits access and critique of these tools. Objectives:  The aim of this study is to investigate how machine learning (ML) and data science can be used to accurately estimate urban-scale air and land-surface temperatures between 1km and 30m resolution, respectively.  A key objective is to develop a model or system that can downscale temperature to these resolutions for any past or future timestamp or day for which we have coarse (background) temperature information from reanalysis or forecast products. Methodology:  We intend to make use of machine learning models to learn the relationship between important climate variables that drive and affect urban-level temperature, and high-resolution urban temperatures.  This includes making use of satellite images, for example from the MODIS, Landsat and Sentinel instruments, in conjunction with coarse temperature fields from reanalysis or forecast products like ERA5 or CFSv2.  Results:  Preliminary results show that machine learning models are capable of using relevant information from satellite images, including information about vegetation, building density and location to accurately characterize the UHI effect.  Detailed analysis of prediction results from a held-out test dataset has been conducted to highlight how well the ML model is able to reproduce the expected spatial and temporal variability. The ML model was trained on approximately 10 years of data with validation during training, accomplished with one year of data. The model was trained on data for a single location, namely, Johannesburg to produce daily maximum near-surface air temperature (T2m) maps at a spatial resolution of 1 km. The ML model predictions, calculated over 1 year of unseen test data, provides an RMSE of <1.8 °C. Conclusion and Next Steps:  Promising initial results indicate that machine learning can be considered a good candidate for accurate downscaling of temperature to urban-scale spatial resolution.  This will allow for large-scale UHI studies that can be used to investigate the effects of extreme temperature on health outcomes, such as with the HE2AT Center.  The initial demonstration study focused on Johannesburg, but this could be extended to other cities in Africa, or even across the globe.  More research will also be conducted on what supplementary predictors and layers can provide more information for better UHI estimation, for example albedo, soil moisture, and digital elevation.  Further research will be done to develop or incorporate novel normalization strategies for the remote sensing data to overcome the limitations, specifically reduced model generalizability, of normalizing such data using the means and standard deviations. All of the data acquisition and pre-processing pipelines will be operationalized in order for the model to be used in real-world use-cases. Ultimately, the goal of this work is to have an operational model capable of predicting high-resolution near-surface air and/or land-surface temperature across space and time. Authors (in surname alphabetical order) Muaaz Bhamjee (1),* Zaheed Gaffoor (1) Tamara Govindasamy (1) Craig Mahlasi (1) Sibusisiwe Makhanya (1) Etienne Vos (1) Affiliations (1) IBM Research Africa, 45 Juta Street, Tshimologong Precinct, Braamfontein, Johannesburg, 2001, South Africa * Presenting Author Presenting Author Bio Dr. Muaaz Bhamjee is currently a Staff Research Scientist at IBM Research - Africa. He has a background in Mechanical Engineering and Applied Mathematics. He was involved in research and published in the fields of Computational Fluid Dynamics (CFD), Heat and Mass Transfer, Experimental Fluid Dynamics, Lattice Boltzmann Modelling, Multiphase Flow, Positron Emission Particle Tracking (PEPT), High Performance Computing (HPC), Solar and Renewable Energy, Solar Air Heating, Atomic Layer Deposition (ALD) Modelling, Fluid-Structure Interaction (FSI), Fluid Dynamics in Biomedical Applications and Engineering Education. The primary focus of his research was on the use of Navier-Stokes and Lattice-Boltzmann based computational fluid dynamics (CFD) approaches in modelling of the multi-scale fluid dynamics, heat and mass transfer processes. Currently he is involved in research related to data-driven and machine learning approaches to Climate. Presenting Author Email Address: Muaaz.Bhamjee@ibm.com