Advanced geo-vision transformer for prediction and mitigation of urban heat islands in semi-arid regions

Urban Heat Islands (UHIs) pose a significant environmental health challenge, especially in semi-arid areas, where growing urbanization exacerbates thermal stress. Traditional studies on UHI often lacked predictive capabilities and in-depth spatial analysis. Therefore, the current study explores the implementation of existing Geospatial Artificial Intelligence (GeoAI) to develop a Geospatial Vision Transformer (GeoViT) model for predicting Land Surface Temperature (LST) at the University of Sharjah (UOS), UAE. The current study develops a predictive GeoViT framework to estimate LST and identify UHI hotspots, generating a high-resolution profile. The main objectives of the analysis are presented as follows: map LST and detect UHI zones using Landsat-8/9 and in situ data and compared for summer and winter, identify the major ones by considering environmental factors along with their weighted contribution to LST, superpixel analysis (an image segmentation technique that combines adjacent pixels into homogeneous region) using Landsat 8/Sentinel-2 to understand the difference with noise reduced LST (LST with minimized uncertainties) and urban features, Analysis of LST along with the LULC trends in UHI intensities, and generate UHI intensity patterns. The model consists of 13 variables, including albedo, sky view factor (SVF), and building heights, to enhance prediction accuracy. The model's predictive accuracy (83%, 85%), measured by precision (0.62, 0.73), recall (0.83, 0.60), and F1-score (0.71, 0.66), demonstrates support for both summer and winter in estimating LST at the University of Sharjah. In addition, the study proposes mitigation strategies based on the spatial patterns and driving factors identified through the Geo-ViT model predictions that explain the optimisation of green areas, the reduction of heat-absorbing surfaces, and the increase in the use of reflective surfaces. The results indicate that the Landsat-8-derived LST estimates (39–56 °C in winter; 48–65 °C in summer) and the model-predicted LST values (45–57 °C in winter; 44–65 °C in summer) align with the in-situ data, showcasing a strong consistency. The predicted UHI intensity for 2025 shows that the most heat-producing areas are airports and bare land, with 72% in summer and 70% in winter.