Predicting Land Surface Temperature Using Meteorological, Environmental, and Urban Morphological Factors

Abstract

Accurate prediction of land surface temperature (LST) at high spatial resolution is crucial for urban planning, climate adaptation, and heat mitigation strategies. This study develops and evaluates machine learning model for predicting LST at different times of the day and across seasons in New York City using ECOSTRESS satellite observations and multiple environmental predictors. We constructed a comprehensive dataset ranging from August 2018 to June 2025 with over 31 million observations at 70-meter resolution, integrating atmospheric reanalysis data (ERA5), vegetation indices (NDVI), topographic features, urban morphological parameters, and solar geometry information. The study implemented comprehensive quality control procedures for ECOSTRESS observations and a hierarchical gap-filling strategy for missing NDVI data to ensure that we have a prediction map with no gaps. Two predictive models, multiple linear regression (LR) and random forest regression, using a total of 42 predictors were evaluated on this dataset. For absolute LST prediction with pixel-wise splitting, RF achieved exceptional performance (R2 = 0.99, RMSE = 1.190C) compared to LR (R2 = 0.89, RMSE = 1.190C). However spatial anomaly prediction revealed fundamental differences as LR explained almost zero spatial variance compared to 64% (R2 = 0.64) spatial variance explained by RF. This demonstrates that urban heat island patterns are driven by nonlinear, interactive processes that LR model fails to capture. Feature importance plot shows that absolute LST is primarily driven by solar/temporal patterns and prevailing meteorology, with air temperature being the most dominant factor. When predicting spatial anomalies, obtained by subtracting the spatial mean for that time period from each image at each time step, a configuration we are referring to as unpooled model -- the importance of urban morphology increased by 17-fold indicating that urban forms have less effect on absolute temperature but works as a critical factor for determining spatial heterogeneity. The RF model was able to maintain robust performance across seasons (R2 ≥ 0.97) and times of day (R2 ≥ 0.90) while LR performance varied from winter R2 of 0.47 to spring R2 of 0.88. Spatial performance mapping revealed that RF consistently outperformed LR across all locations. R2, RMSE, and MAE are higher in less densely urbanized areas such as Staten Island and outer Queens, and low in densely urbanized areas such as Manhattan and central Brooklyn. By analyzing persistent summer hotspots defined as locations consistently warmer than the 95th percentile in summer months, we found that RF underpredicts these areas by 0.530 C on average. To validate the findings, we looked at R2 for different feature configurations (10 to 42 predictors). Sensitivity analysis using 10-fold cross-validation using 200,000 observations showed that spatial anomaly, or unpooled model prediction performance (R2 = 0.32) is more sensitive to evaluation methodology than the pooled models. Summer afternoons showed the highest prediction errors for spatial anomalies (RMSE = 2.980C). These findings demonstrate that pixel-wise validation is good for gap-filling but we still need separate analysis of spatial anomalies separate from absolute LST. In future work, we need to conduct temporal validation in addition to the pixel-wise validation we have done here.