LCZ-Stratified Estimation of Near-Surface Air Temperature from Land Surface Temperature: Evidence from Central Xi’an, China

Near-surface air temperature (Ta) is essential for urban heat assessment, but ground observations are spatially sparse. This study combined ECOSTRESS land surface temperature (LST), concurrent Ta and built-environment (BE) indicators in four local climate zone (LCZ) categories in central Xi’an. The analysis comprised 176 site-overpass records from 123 sites across six events. Under ten repetitions of site-grouped five-fold cross-validation, LST-only CatBoost yielded RMSEs of 2.69, 1.63, 4.41 and 2.22 °C in LCZ5-II, LCZ4, LCZ2 and LCZ6-II, respectively. OLS performed better in LCZ4 (1.41 °C). Adding 14 screened BE variables increased the equal-weight mean RMSE across LCZs from 2.74 to 3.29 °C; this disadvantage persisted across the tested model capacity settings in LCZ5-II, LCZ4 and LCZ6-II. In event-held-out tests, the LST-only RMSE rose to 5.84, 6.57 and 7.12 °C in those three LCZs. The dominant LCZ2 event lacked sufficient training support for evaluation. LCZ2’s inverse prediction–observation association arose predominantly between event means, and all tested site-grouped CatBoost configurations retained negative predictive R2 values. Within this small sample, additional BE inputs offered limited predictive value, and errors increased substantially when transferring to new events.