A novel approach to estimating urban land surface temperature by the combination of geographically weighted regression and deep neural network models

Growing concerns on excessive urban heat call for better approaches to modeling urban thermal environment and developing effective mitigation strategies. A hybrid model integrating the geographically weighted regression (GWR) and deep neural network (DNN) was developed to estimate land surface temperature (LST). This model was compared with three other data-driven approaches to predicting LST, including the ordinary least squares (OLS) regression, GWR, and DNN. Sixteen satellite image datasets (a total of 155,728 data points) during a four-year period in Hong Kong were used for model development, validation, and comparison. The datasets cover two distinguishable geographical regions and consist of sixteen explanatory variables from five groups, including (1) land use and land cover (LULC) composition and surface characteristics, (2) LULC configuration, (3) urban form, (4) anthropogenic activities, and (5) location and local climate. The results indicate that the hybrid model performs the best in terms of model fitness and prediction accuracy, with R2 equal to 0.85 and 0.73 and the mean squared error (MSE) equal to 0.52 and 0.70 in the two regions, respectively. Compared to the OLS, DNN, and GWR models, the overall R2 for all the datasets of the hybrid model increases by 97.3%, 16.6%, and 6.9%, respectively. The promising result of the hybrid model is due to its ability to capture both spatial heterogeneity and address possible correlations between explanatory variables. Sensitivity of LST to various explanatory variables is also discussed and strategies to mitigate excessive heat are recommended. This study is anticipated to contribute to model development in urban LST estimation and quantitative evaluation of those factors driving LST variations.