Downscaled high spatial resolution images from automated machine learning for assessment of urban structure effects on land surface temperatures

Urbanization has profoundly reshaped urban morphology and land cover while degrading the thermal environment. Despite numerous studies exploring correlations between two-dimensional (2D) and three-dimensional (3D) urban features and land surface temperatures (LSTs), understanding the impact of urban structural effects on LSTs remains unclear due to limited high-spatial-resolution satellite data. This study addresses this gap by integrating satellite images and volunteered geographical data, employing automated machine learning through Autokeras to downscale LSTs to a 10-m spatial resolution. Subsequently, a stepwise regression model quantified the relationships between various urban feature indicators and LSTs within urban blocks. Results indicated the Autokeras-trained LST-prediction model achieved high accuracy (RMSE: 0.528 K, MAE: 0.317 K, R2: 0.973), demonstrating its efficacy in generating accurate 10-m LSTs from SDGSAT-1 satellites. The stepwise regression model effectively characterized relationships between urban features and LSTs, yielding RMSE, MAE and R2 of 1.142 K, 0.881 K and 0.646, respectively. LSTs exhibited heighted sensitivity to albedo, emissivity, normalized difference vegetation index, building height, and ratio resident-area index, with their combined weights exceeding 70 %. Comparisons with SDGSAT-1 raw data and Landsat 8, which operates at a lower spatial resolution (30 m), underscored the finer delineation capabilities of our high-resolution LSTs across heterogeneous land covers. Furthermore, 10-m LSTs showed 4.2 % greater sensitivity to building height than 30-m LSTs, highlighting their ability to better capture cooling effects from 3D structure shadows. This study thus underscores the utility of high-resolution LST data in urban planning and climate adaptation strategies. •A novel method to downscale urban LSTs to 10 m via automated machine learning.•LSTs with 10-m resolution precisely delineate edges of land covers via temperatures.•Lower-resolution data overestimate LSTs in dense building areas by about 1–4 K.•Higher-resolution LSTs increase building height's weight by 4.2 %.