Enhancing Spatial Resolution and Accuracy of Land Surface Temperature: Integration of Regression-based and Surface Energy Balance Models

Land surface temperature (LST) derived from satellite thermal sensors is a crucial dataset for environmental and urban studies. However, the limited spatial resolution and accuracy of these datasets may present significant challenges for various applications. This study introduces two innovative approaches to improve the spatial resolution and accuracy of LST: (1) a regression-based model integrating multiple sources of information and (2) a physically-based model of the Surface Energy Balance (SEB). The regression-based model employs, a decision-level fusion approach to minimize the impact of various error sources. Regression approaches include various combinations of regression models and different training and implementation strategies. In this study, four models were employed to develop an appropriate relationship between LST and environmental parameters: (1) Partial Least Squares Regression (PLSR), (2) Support Vector Regression (SVR), (3) Artificial Neural Networks (ANN), and (4) Random Forest Regression (RFR). For different model training and implementation approaches, the following strategies were considered: (1) Global Window Strategy (GWS), (2) Conceptual Window Strategy (CWS), (3) Regular Moving Window Strategy (RLWS), (4) Object-Based Window Strategy, and (5) Decision-Level Integration Window Strategy (DIWS). The second approach presents a novel physical model for enhancing the spatial resolution of LST using energy balance equations across different land cover types. For the first time, this model combines the Temperature Separation Principle (TSP) and Thermal Unmixing Model (TUM) frameworks to improve accuracy. This integration ensures that the physical nature of the spatial resolution enhancement process significantly mitigates scaling effects on LST accuracy, maintaining or improving the absolute accuracy of LST. The study uses diverse datasets, including imagery from Landsat 8 and MODIS Terra satellites, land cover maps, impervious surface percentages, digital elevation models, building heights, population density, and ground-based measurements. The study area included six cities in the United States (Chicago, Dallas, Minneapolis, Phoenix, Seattle, and Kansas), 13 cities in Europe (Lisbon, Madrid, Zamora, Bucharest, Vienna, Prague, Paris, London, Warsaw, Copenhagen, Herning, Stockholm, and Helsinki), and one city in Iran (Tehran). The findings reveal that in urban and agricultural areas, biophysical characteristics predominantly influence LST distribution, whereas topographical features have a greater impact in mountainous regions. Urban areas exhibit stronger effects of surface texture and neighborhood characteristics on LST distribution compared to other regions. Incorporating neighborhood effects and landscape parameters in the spatial resolution enhancement process reduced the LST error by 0.8 K in warm seasons and 0.4 K in cold seasons. Furthermore, improving the spatial resolution of LST from 1000 m to 30 m using the regression-based model at the decision-making level and the SEB model reduced the LST error by an average of 2.5 K (3.4 K) in warm seasons and 1.2 K (1.8 K) in cold seasons. The SEB model also provided additional insights into temperature distribution by accounting for evapotranspiration and energy fluxes. These findings underscore the high potential of the proposed approaches in simultaneously improving the spatial resolution and accuracy of LST, making them highly applicable for environmental and urban studies. 

Keywords: LST, Spatial Resolution Enhancement, Surface Energy Balance, Regression Models, Decision-Level Integration