Deep Learning–Based Modeling of Surface Urban Heat Island Integrating GNSS-Derived Atmospheric Humidity and Multi-Source Urban Data

Surface Heat Island (SUHI) is a result of complex and non-linear interactions between atmospheric processes and urban surface features. These interactions operate at different spatiotemporal scales. Research shows that surface coverage and urban morphology affect the urban thermal environment; however, most SUHI modeling approaches still rely on surface features and mostly ignore important parameters such as atmospheric humidity and precipitation. This problem limits the ability of existing SUHI models to accurately represent interactions between the surface and atmosphere and thermal variability.

This research presents a deep learning-based framework for SUHI modeling which is developed based on integrating urban, atmospheric, and environmental features. The proposed framework integrates Landsat-derived land surface indicators, including Land Surface Temperature (LST), the Normalized Difference Vegetation Index (NDVI), which represents vegetation cover, the Normalized Difference Built-up Index (NDBI) and Night Time Light (NTL), which  characterize built-up areas, and the Normalized Difference Water Index (NDWI), which represents surface water bodies, with GNSS-derived Precipitable Water Vapor (PWV) as a measure of atmospheric humidity and Global Precipitation Measurement (GPM) data. Other effective parameters include topography from the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) digital elevation model and population density in different part of the city.

A Convolutional Neural Network (CNN) architecture is developed to capture spatial dependencies in urban areas and to understand the non-linear interactions between surface, atmospheric, social, and environmental features. This model is composed of many stacked convolutional layers with regularization and pooling algorithms to maintain generalization and preserving spatial structure. SUHI intensity, defined as the contrast between LST in urban and rural areas, is used as the target for prediction. Model training and validation are based on cross-validation to assess robustness and transferability across different temporal subsets.

The case study is Wrocław, Poland that has been experiencing rapid urban development and has undergone substantial land-use and structural transformation over the past decade. Comparisons between results from models that include and exclude humidity and precipitation demonstrate that GNSS-derived PWV and precipitation play a significant role in SUHI modeling.

The results highlight the importance of accounting for urban–atmosphere interactions in SUHI modeling. This deep learning framework provides a practical basis for subsequent eXplainable Artificial Intelligence (XAI) analyses. XAI analysis can identify SUHI drivers and support climate-resilient urban planning and heat mitigation strategies.