Data-Driven Evaluation of Climate Projections Using Statistical and Deep-Learning Methods for Pavement ME Design

Abstract

Environmental conditions have a major influence on pavement design, performance, and long-term service life. In Oklahoma state, diverse and often extreme climatic conditions can greatly accelerate deterioration through cracking, rutting, moisture damage, and surface wear. To build more resilient transportation infrastructure, it is critical to integrate reliable climate projections into the mechanistic–empirical (ME) pavement design process. This study evaluates and ranks 10 of NASA’s NEX-GDDP-CMIP6 for Oklahoma under three shared socioeconomic pathways: SSP1-2.6 (low emissions/strong mitigation), SSP2-4.5 (intermediate stabilization/moderate mitigation), and SSP5-8.5 (high emissions/limited mitigation). Two complementary approaches were employed to rank model suitability. The first used traditional statistical metrics aggregated with the technique for order preference by similarity to ideal solution multi-decision-making method to assess how closely models reproduced observed temperature and precipitation. The second applied a Siamese long short-term memory neural network to evaluate temporal similarity between observed and simulated time series, capturing nonlinear patterns and seasonal dynamics that standard statistics may overlook. Together, these approaches address the limitations of using statistical metrics alone, which can be sensitive to daily variability or extreme values, by balancing accuracy with temporal consistency. Results indicated that the SSP5-8.5–CanESM5 model provided the strongest overall performance, followed by SSP2-4.5–FGOALS and SSP5-8.5–FGOALS. The framework not only identifies the most suitable models for pavement design in Oklahoma but also demonstrates a transferable methodology for climate-adaptive infrastructure planning in other regions.