Reconstructing Urban Surface Temperatures: A Machine Learning Approach to Bridging Temporal Gaps in High-Resolution Data

Urban heat islands and extreme heat events are intensifying due to climate change, especially in densely built environments. Remote sensing of land surface temperatures (LST) offers valuable insights for analyzing and mitigating urban heat risks. However, a major limitation of satellite-derived LST data is the trade-off between spatial and temporal resolution. High-resolution products such as those from Landsat provide fine spatial detail but suffer from low temporal coverage, limiting their usefulness for time-critical analyses.

In this study, multiple machine learning approaches are presented to reconstruct high-resolution urban LST data in sub-daily time steps by bridging temporal gaps using observations from the ECOSTRESS sensor on board the ISS. Using Madrid as a case study, random forest, gradient boosting, and artificial neural network models were trained on ECOSTRESS LST data together with a comprehensive set of explanatory variables, including local weather and radiation measurements, ERA5 reanalysis data, and Sentinel-2 surface reflectance indices.

Results show that the different model architectures exhibit varying strengths and weaknesses. The precision of the reconstructions varies with land use; urban areas tend to be reconstructed more accurately than non-built-up, sparsely vegetated areas. Comparing each model’s strengths and weaknesses highlights the potential use of data-driven methods to overcome observational limitations and generate continuous, high-resolution thermal datasets across the diurnal cycle.

By investigating the use of machine learning techniques for the reconstruction of Madrid’s land surface temperature, this work shows a potential pathway to overcome data gaps in high-resolution data on a broader scale. Therefore, it contributes a step toward continuous land surface temperature data, which may help improve the understanding of local heat waves and possible adaptation strategies.