Urbanization combined with climate change is intensifying urban thermal stress, requiring proactive and spatially explicit heat-risk assessment. This study develops a scenario-based framework that integrates machine-learning–enhanced land cover simulation with land surface temperature (LST) prediction to compare future thermal outcomes under Urban Growth (UG) and Ecological Conservation (EC) scenarios through 2034. Within the FLUS framework, probability-estimation modules were empirically evaluated, and XGBoost achieved the most stable predictive performance; validation against the observed 2024 land cover map confirmed simulation reliability. By 2034, the EC scenario reduced the study-area mean LST by 0.10°C relative to UG, while localized cooling within the existing urbanized area reached 0.6°C. Distance-based analysis showed that mitigation effects were concentrated within approximately 300 m of the urban boundary. When high-temperature areas were defined as pixels exceeding 32.10°C (top 10% threshold), their extent reached 90.64 km² under UG but was constrained to 77.77 km² under EC, representing a reduction of 12.87 km². These findings demonstrate that conservation-oriented land-use strategies can meaningfully influence both average warming levels and the spatial redistribution of urban heat risk.