Urban heat islands (UHI) pose significant challenges in arid cities, where limited water availability constrains the effectiveness of vegetation-based cooling strategies. This study examines the relationships between land surface temperature (LST) and urban form across four urban functional zones (UFZs) in Urumqi, a representative arid city in Northwest China. Multi-source remote sensing data were used to derive two-dimensional (2D) landscape metrics and three-dimensional (3D) morphological indicators for 1039 samples. Zone-specific XGBoost models were developed, and SHapley Additive exPlanations (SHAP) were applied to analyze feature contributions. The models show moderate predictive performance, with test R2 values ranging from 0.354 to 0.612. Among all variables, the sky view factor (SVF) is consistently identified as the most influential predictor, indicating a strong association between shading-related urban morphology and LST. In contrast, landscape composition variables exhibit more varying patterns across UFZs. Nonlinear relationships are observed for several variables. For example, PLAND_g shows threshold-like and zone-dependent responses around approximately 15%-20%, particularly in residential and public service zones, whereas such patterns are less evident in industrial areas. Impervious surface proportion and building height also show diminishing effects at higher values. These findings highlight the importance of considering both 3D urban form and functional zoning in arid environments, and provide insights for more context-specific urban heat mitigation strategies.