Accelerated Arctic warming is destabilizing permafrost, threatening ecosystems, infrastructure, and northern communities, yet permafrost thermal dynamics remain poorly characterized and highly uncertain due to sparse observations and limited representation of fine-scale heterogeneity. Here we combine kilometer-scale distributed temperature sensing (DTS) with data-driven hysteresis modeling to resolve spatiotemporal variability in permafrost temperatures across disturbed and undisturbed Arctic landscapes near Utqiaġvik, Alaska. A 2-km fiber-optic array recorded continuous ground temperatures from 2021 to 2024, revealing persistent warming associated with civil infrastructure and pronounced thermal heterogeneity in patterned tundra. Ice-wedge polygon troughs consistently exhibit lower temperatures than polygon centers, highlighting the role of subsurface ice distribution in controlling ground thermal regimes. Using these observations, we develop a multivariate hysteresis model that captures lagged ground–air temperature responses and incorporates snow, precipitation, wind, atmospheric pressure, and ground surface conditions. The model accurately reproduces observed permafrost temperatures, fills observational gaps, and enables projections under future climate scenarios. Projections indicate continued warming through 2075, with enhanced temperature increases in infrastructure areas and ice-rich terrains. Our results demonstrate the power of DTS to resolve permafrost thermal heterogeneity at unprecedented scales and provide a transferable framework for predicting permafrost temperature dynamics in a rapidly warming Arctic.