A real-time characterizing, forecasting and mapping framework of extreme pavement temperature hazards in transitional climates: an operational case study

Extreme pavement temperature (ETs) threatens road infrastructure through deformation, accidents, and black ice. This study analyzed 2015–2018 minute-resolution data from seven highway stations in China's Huaihe River Basin and proposed an analysis–modelling–mapping framework that combines spatiotemporal characterization of ETs, deep learning and spatial interpolation. Three forecasting challenges emerged: significant data imbalance (ETs/not-ETs ratio < 0.32; hot/cold extremes ratio < 0.83); spatial heterogeneity non-correlated with geographical positions; and divergent durations (hot: < 6 h, occasionally < 2 h in September; cold: > 12 h, even exceeding 24 h in January). Specifically, a multi-scale temporal fusion forecasting network (MSTF-ROAD) featuring multi-scale temporal feature pyramids, dual-pooling mechanisms (Avg–Max pooling), real-time attention and conventional events downsampling was proposed. Evaluations demonstrated MSTF-ROAD's superiority with 0.67 °C MAE, capturing >93% cold extremes and >74% hot extremes, effective trends/inflection tracking. The dynamic geospatial hazard maps with lightweight ETs prediction architectures attained 100-m spatial resolution and enabled minute-level temporal synchronization. The primary innovation of the MSTF-ROAD framework lies in its theory-guided lightweight model architecture, which collectively overcomes the critical challenges of data imbalance and spatiotemporal heterogeneity, achieving a balanced and accurate prediction of both hot and cold extremes. This interdisciplinary methodology directly supports Sustainable Development Goals 9 and 13.