Forecasting Urban Heat Islands: A Neural Network Approach Using Remote Sensing Data

Urban heat islands significantly challenge environmental sustainability and public health, creating localized areas within cities with higher temperatures. Addressing these issues requires predictive tools for precise temperature forecasts to aid urban planning and policy decisions. Although satellite-based land surface temperature (LST) monitoring has potential, data from the ESA Copernicus Sentinel-3 mission face two key limitations: inadequate spatial resolution for urban-scale differentiation (1 km per pixel bidaily LST measurements) and the disparity between land surface and air temperatures.

This research introduces a machine learning model designed to predict maximum daily air temperatures at a spatial resolution of 20 meters per pixel, sufficient for the recognition of temperature differences between individual city blocks. For each day the inference is run, the model produces a seven-day temperature forecast. Our technology utilizes a visual transformer-based architecture, which distinguishes itself by being more compact and computationally efficient than traditional convolutional neural networks (CNNs), achieving a mean absolute error (MAE) of 2°C across seven-day temperature predictions for three major European cities.

The model uses multiple remote sensing and weather forecast data. The first input is LST data fromSentinel-3. It also uses NDVI data from Sentinel-2, sensitive to vegetation health and density. Meteorological data include forecasted temperature, pressure, humidity, wind, and more. For topographic data, two sources are used: the Digital Elevation Model for terrain altitude and the Copernicus Urban Atlas for land use classification. All input data is resized to the required dimensions and combined into a single 3D tensor for the model. Circular encoding is used to incorporate the day of the year and time of day of the Sentinel-3 passage. All inputs, except for the weather data, are stacked and combined with the weather data for the predicted day, then passed to the model. This process is repeated for each of the seven days to generate the temperature predictions.

 

Temperature measurements used for target for ML training are sourced from on ground stations and processed into a 2D matrix, with pixel values showing the average maximum temperature recorded by each station within the pixel's area. Pixels with no active stations are marked as invalid. For each valid pixel, the mean squared error (MSE) loss between the model's predicted temperature and the ground truth is computed to update the model weights. An encoder-decoder architecture is used to translate these multidimensional inputs into a set of two-dimensional temperature maps. The chosen encoder is a Mixed Transformer model (MiT), and the decoder is a simple cascade of convolution-upsample.

The model is embedded in a continuous pipeline for uninterrupted operation. Its daily workflow automatically retrieves data, preprocesses it, and generates temperature mappings. Seven-day temperature forecasts are uploaded to a dashboard, presenting predictions as overlays on urban landscapes. This solution is part of UP2030, a project supported by the EU's Horizon Europe program, which guides cities through socio-technical transitions towards climate neutrality.