A framework to assess permafrost thaw threat for land transportation infrastructure in northern Canada

The Arctic region, despite its small population, contributes significantly to the global GDP. Recent decades have seen increased socioeconomic development, with projections for further population growth and increased marine transportation due to sea ice loss. However, climate warming destabilizes permafrost, posing a significant threat to Arctic communities and infrastructure. Permafrost, ground remaining below 0°C for at least two years, covers a substantial portion of the Northern Hemisphere, supporting millions of inhabitants and extensive infrastructure, including roads, railways, airports, and pipelines. Arctic warming exceeds global averages, increasing permafrost thaw sensitivity. Thawing leads to ground deformations and instabilities, impacting infrastructure accessibility and design. Several studies highlight the substantial regional impacts of permafrost degradation on communities and infrastructure, emphasizing the need for adaptation measures. While near-surface air temperature is often used as a threat indicator, it does not fully capture the complexities of the ground thermal regime. Numerical simulations are challenging due to the complexities of the surface energy budget. This study addresses the challenges in permafrost thaw threat assessments using a data-driven framework that leverages reanalysis data (ERA5-Land) and regional climate projections (CanRCM) to predict long-term ground surface temperatures and calculate a thaw index. This index considers both ground surface temperature projections and the present distribution of ground ice for rapid assessment of permafrost thaw threats at regional and subregional scales. The framework is evaluated by applying it to three major land transportation infrastructure projects in northern Canada.

Numerous studies have assessed the regional impacts of permafrost degradation. Estimates suggest significant permafrost degradation by 2050 affecting millions of residents and most Arctic infrastructure. Vulnerability indices for permafrost have been defined using regional ground ice and soil type distributions. Near-surface air temperature is often used as a proxy for permafrost stability due to its correlation with ground temperature. However, the ground thermal regime is complex, influenced by surface energy budget components (ambient temperature, solar radiation, wind, snow cover, vegetation) and thermal properties of the ground. Numerical simulations can model the ground thermal regime, but their high-resolution multi-physics applications at regional scales are infeasible. Data-driven approaches using machine learning (ML), like artificial neural networks (ANN), have been used for short-term ground temperature estimations, but often neglect the sequential nature of climate data and long-term climate change uncertainties. Recurrent neural networks (RNNs), such as LSTM and GRU, address temporal dynamics in time-series data and have shown promise in predicting various climate variables. The spatiotemporal scarcity of climate measurements is addressed using data assimilation and reanalysis models, such as MERRA2 and the ERA family, with ERA5-Land offering high-resolution land surface climate data. However, these reanalysis products might have biases, requiring careful evaluation.

This study develops a framework to assess permafrost thaw threats by predicting long-term ground surface temperatures using ERA5-Land reanalysis data and CanRCM regional climate projections. The framework calculates a thaw index based on projected ground surface temperatures and existing ground ice distributions. Three case studies—Hudson Bay Railway, Mackenzie Northern Railway, and Inuvik-Tuktoyaktuk Highway—are used to evaluate the framework's performance. Model validation is performed using a blind validation scheme, comparing predicted ground surface temperatures to ERA5-Land data over a test period excluded from the training data. Statistical metrics (bias, MAE, RMSE, MaAE) are used to evaluate model accuracy. Ground ice indices are derived from the Ground Ice Map of Canada (GIMC), classifying ground ice into segregated, wedge, and relict ice, with abundance quantified from 0 (no ice) to 1 (high abundance). The permafrost thaw threat index (I_t) is defined as I_t = f(MAGST)(β_sI_s + β_wI_w + β_rI_r), where f is a non-linear thaw criterion function (1/[1 + e^-MAGST]), MAGST is mean annual ground surface temperature, β_s,w,r are weight factors (assumed equal), and I_s,w,r are ground ice abundance indices. The LSTM network, a type of RNN, is employed to predict ground surface temperatures. The models are trained on ERA5-Land data (ambient temperature, solar radiation, wind speed, and snow depth as independent variables, and ground surface temperature as the dependent variable) for each study site on a 0.1° x 0.1° grid. CanRCM projections under RCP 4.5 (moderate) and RCP 8.5 (extreme) scenarios are used to predict future ground surface temperatures. The thaw index is then calculated for the present, mid-century, and end of the century. The methodology utilizes spatial analysis in GIS software to integrate ground ice data with the model outputs to create spatial distributions of the thaw index.

Model validation shows good accuracy in predicting ground surface temperatures, with mean RMSE less than 0.3 °C and MaAE less than 2 °C. Ground surface temperature projections indicate warming in all three study areas under both RCP 4.5 and RCP 8.5 scenarios. The Mackenzie Delta region (Inuvik-Tuktoyaktuk Highway and Dempster Highway) shows the highest warming rates, possibly due to Arctic amplification. The thaw index displays spatial variation largely following ground ice content. The Hudson Bay Railway shows a gradual northeastward increase in the thaw index over time, with the southern section already experiencing thawing. The Mackenzie Northern Railway shows little change in the thaw index despite projected warming, suggesting existing instability. The Mackenzie Delta shows the highest thaw indices, with significant increases projected by the end of the century. The analysis indicates that all three study areas are already under threat, with the northern sections of the Hudson Bay Railway and the Mackenzie Delta facing increasing threats by the end of the century. A summary table clarifies how present and future thaw index values are interpreted in terms of risk.

The developed thaw index provides a rapid assessment of permafrost thaw threat by combining climate warming projections with ground ice distributions, addressing limitations of traditional methods like the n-factor method. The framework's scalability allows for using various projection datasets without model retraining. The framework's long-term ground surface temperature predictions can simplify complex thermal and thermo-hydro-mechanical simulations by providing boundary conditions, addressing the shortcomings of the n-factors method. Additional independent variables, if available, can improve local prediction accuracy. The thaw index formulation can be further refined to include other factors like subsurface conditions and soil properties. The study's findings are relevant for engineering design, infrastructure management, and environmental stewardship. The thaw index highlights areas currently at risk and those facing future threats, informing decision-making related to infrastructure development, maintenance, and adaptation measures.

This study presents a novel framework for assessing permafrost thaw threats to infrastructure, addressing limitations of existing methods. The data-driven approach, using LSTM models, accurately predicts ground surface temperatures and incorporates ground ice distributions to create a comprehensive thaw index. Results highlight existing and future thawing threats, emphasizing the need for adaptation strategies. Future research could focus on refining the thaw index formulation, incorporating more local-scale data, and exploring advanced machine learning techniques to enhance predictive capabilities.

The study's accuracy depends on the quality and availability of input data (ERA5-Land, CanRCM, GIMC). While the model shows good accuracy, inherent uncertainties in the projection datasets influence the results. The equal weighting of different ground ice types might not fully reflect the varying hazards they pose. The study focuses on three transportation corridors, and the findings may not be fully generalizable to all areas with permafrost. Future work could explore different ML models and consider an ensemble of climate projections to account for uncertainties.