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

The Arctic region, though sparsely populated, contributes substantially to global GDP through oil, mining, food, and tourism and is undergoing socioeconomic changes, including increased marine traffic due to sea ice loss. Concurrently, climate warming is destabilizing permafrost across the Arctic and subarctic, threatening communities and extensive transportation infrastructure built on frozen ground. Permafrost—ground that remains at or below 0 °C for at least two consecutive years—covers more than 15% of the Northern Hemisphere and half of the Arctic landmass. Observations show Arctic warming exceeds global averages; permafrost temperatures increased by about 0.39 ± 0.15 °C (continuous) and 0.20 ± 0.10 °C (discontinuous) between 2007 and 2016, and permafrost sensitivity is estimated at 4.0 million km² thawed per 1 °C global mean warming. From a geotechnical perspective, warming reduces soil strength, causing creep, thaw settlement, thermokarst, and landslides, with severe implications for existing and planned land transportation assets. Robust, spatially extensive indicators of permafrost stability are needed for regional and subregional risk assessments, yet data scarcity, complex surface energy processes, and climate uncertainties hinder conventional modeling. This study addresses these challenges by developing a machine learning-based framework to forecast ground surface temperatures and combine them with present ground ice distributions to assess thaw threat for three key Canadian transportation corridors.

Traditional threat indicators often use near-surface air temperature due to its availability in climate models, but the ground thermal regime is governed by the full surface energy budget (radiation, wind, snow, vegetation) and subsurface properties. Numerical multi-physics simulations can estimate ground thermal states, yet they are challenging at regional scales due to complexity and data requirements. Ground surface temperature (GST) better reflects the surface energy budget than air temperature but is not widely measured and is often inferred from air temperature via n-factors, which do not capture nonstationary climate or multi-variate energy exchanges. Data assimilation and reanalyses (e.g., ERA5, ERA5-Land, MERRA2) help address spatiotemporal data gaps; ERA5-Land provides land surface variables since 1950 at ~9 km resolution and performs well in many regions, with known biases in circumpolar permafrost areas. Machine learning has been used for short-term ground temperature forecast (ANN, ELM, GRNN, BPNN, RF, ARIMA), but many approaches ignore temporal dependencies. Recurrent neural networks, notably LSTM and GRU, capture sequential dynamics and have been applied to climate and energy variables. Prior work showed station records and regional climate projections can forecast long-term GST, but scaling requires spatially continuous input data. This study leverages ERA5-Land and CanRCM projections with LSTM to produce spatially resolved long-term GST forecasts and integrates them with mapped ground ice to quantify thaw threat.

The framework estimates permafrost thaw threat via a thaw index computed from projected mean annual ground surface temperature (MAGST) and present ground ice abundance. 1) Data sources: ERA5-Land reanalysis provided labeled daily training data (dependent variable: GST; independent variables: near-surface air temperature, shortwave and longwave downward radiation, zonal and meridional wind speed, and snow depth) on a 0.1° × 0.1° grid from 1950–2022 (>26,000 daily entries per node). CanRCM regional climate projections provided daily independent variables under RCP 4.5 and RCP 8.5 to 2100. Ground ice distributions were derived from the Ground Ice Map of Canada (GIMC) for the top 5 m of permafrost, categorized into segregated, wedge, and relict ice. 2) Model architecture and training: Classic LSTM models were trained independently per grid node. Dataset split was 75% training and 25% testing; k-fold analysis was used to monitor performance on unseen data. Key hyperparameters (per node): epochs 50, batch size 100, input features 6, sequence length 7 days, 1 LSTM layer with 50 hidden units. After training on ERA5-Land, CanRCM predictors were input to forecast daily GST through 2100; MAGST was then computed for past (1950–1960), present (2010–2020), mid-century (2040–2050), and late-century (2090–2100). 3) Thaw index: The permafrost thaw threat index It is defined as It = f(MAGST) × (βs Is + βw Iw + βr Ir), where Is, Iw, Ir are normalized ground ice abundance indices (0–1) for segregated, wedge, and relict ice, respectively, mapped from GIMC qualitative classes (none, negligible, low, medium, high) to quantitative values (0, 0.25, 0.5, 0.75, 1). Equal weights were used (βs = βw = βr = 0.33). The thaw criterion f(MAGST) is a nonlinear logistic function reflecting thermal offset between frozen and thawed states, increasing with MAGST and centered near 0 °C. 4) Study areas: Three transportation corridors in northern Canada were assessed: a 400 km section of the Hudson Bay Railway (Thompson to Churchill), the full 969 km Mackenzie Northern Railway (Grimshaw to Hay River), and a 324 km section encompassing the Inuvik–Tuktoyaktuk Highway and part of the Dempster Highway (Fort McPherson to Tuktoyaktuk). 5) Validation: Blind validation compared predicted GST to ERA5-Land GST over an out-of-sample test period (2017–2023) using bias, mean absolute error (MAE), root mean square error (RMSE), and maximum absolute error (MaAE) at each grid node. Projection datasets (CanRCM) were not validated as ground truth is unavailable and model outputs carry inherent uncertainties. 6) Outputs: Spatial MAGST and thaw index maps and time series at points of interest for present, mid-century, and end-century under RCP 4.5 and RCP 8.5.

- Model performance (2017–2023 test period, spatial means and maxima): Hudson Bay Railway: mean bias error 0.38 °C (max 1.40), MAE 0.44 °C (max 1.40), MaAE 0.72 °C (max 1.96), RMSE 0.28 °C (max 2.05). Mackenzie Northern Railway: mean bias 0.23 °C (max 1.05), MAE 0.34 °C (max 1.05), MaAE 0.65 °C (max 1.58), RMSE 0.19 °C (max 1.20). Inuvik–Tuktoyaktuk Highway: mean bias 0.13 °C (max 1.33), MAE 0.39 °C (max 1.33), MaAE 0.80 °C (max 2.03), RMSE 0.24 °C (max 1.94). - Ground surface warming: By 2100, MAGST at major points of interest increases by 1.03–2.90 °C under RCP 4.5 and 3.48–4.82 °C under RCP 8.5, with greater warming at the Mackenzie Delta consistent with Arctic amplification. Averaged across sites, warming is about 2.1 °C (RCP 4.5) and 4.1 °C (RCP 8.5) by century’s end. - Thaw threat spatial patterns: Hudson Bay Railway shows a northeastward increase in thaw index aligned with higher ground ice content; minor increases over time at Churchill and Herchmer, with an abrupt rise between Herchmer and M’Clintock indicating a shifting thaw front toward the northeast. - Mackenzie Northern Railway: Despite projected warming, the thaw index remains largely unchanged across time horizons, implying thaw processes have already commenced in many segments; south of Paddle Prairie, thaw index is near zero due to limited excess ground ice. - Mackenzie Delta (Inuvik–Tuktoyaktuk/Dempster Highways): Highest thaw indices among study areas, with clear temporal increases. In Tuktoyaktuk, thaw index rises from ~0.15 (2010–2020) to ~0.34 (RCP 4.5) and ~0.62 (RCP 8.5) by 2090–2100. Strong spatial gradients reflect abrupt changes in ground ice content. Coastal jaggedness arises from coastline discrepancies between ERA5-Land and CanRCM grids. - Overall assessment: All study regions are currently under thaw threat; northern sections of the Hudson Bay Railway and much of the Mackenzie Delta will face increasing threat by late century.

The thaw index contextualizes risk relative to present conditions, enabling stakeholders to gauge when and where instability may occur over infrastructure design lifetimes. An increasing thaw index through time indicates present MAGST near or below freezing transitioning to above freezing, elevating future risk, with magnitude modulated by ground ice abundance. Conversely, unchanged but nonzero indices suggest ongoing thaw risk, possibly reflecting already above-freezing MAGST or ongoing degradation. Application to the three corridors reveals dynamic spatial and temporal threat patterns: a migrating thaw front along the Hudson Bay Railway, early-stage or ongoing thaw along the Mackenzie Northern Railway, and strong, increasing threats across the Mackenzie Delta. The framework’s performance (mean RMSE < 0.3 °C, MaAE < 2 °C) supports regional and subregional applications. Beyond rapid screening, the predicted long-term GST can serve as boundary conditions in detailed thermo-hydro-mechanical models, improving upon static n-factor approaches by accounting for climate nonstationarity and multivariate surface energy processes. Incorporating additional local layers (e.g., snow density, vegetation, soil thermal properties, topography, taliks, drainage) could refine local assessments, provided such variables are available in both training and projection datasets. Weighting different ground ice types to reflect distinct hazards (e.g., retrogressive thaw slumps in relict ice versus frost heave and settlement with segregated ice) can tailor the index to specific infrastructure concerns.

The study presents a scalable AI-based framework that combines LSTM-predicted ground surface temperatures with mapped ground ice to rapidly assess permafrost thaw threat over regional and subregional scales. It addresses key challenges of data scarcity, complex surface energy budgets, and the need for practical, spatially continuous indicators. Validation demonstrates strong agreement with reanalysis-based GST, and application to three northern Canadian transportation corridors indicates present-day thaw threats and increasing risks by century’s end, particularly in the Mackenzie Delta and northern Hudson Bay Railway. The framework can support engineering design, asset management, and environmental stewardship by informing site selection, maintenance prioritization, and adaptation planning. Future work should integrate higher-fidelity training data, additional local information layers, ensemble climate projections, refined ground ice weighting schemes, and more advanced sequence learning techniques to reduce uncertainty and enhance local relevance.

- Predictions inherit uncertainties from training and projection datasets (ERA5-Land, CanRCM, and GIMC), including known GST biases in reanalyses and coastal grid inconsistencies causing null predictions along some shorelines. - Validation was limited to the ERA5-Land-based test period (2017–2023); projections could not be validated due to lack of ground-truth dependent variables and inherent climate model uncertainties. - The thaw index uses equal weights for segregated, wedge, and relict ice; differing hazard potentials across ice types are not explicitly represented. - Active layer thickness changes are not directly modeled in the index due to higher uncertainty at regional scales; the approach relies on MAGST as a proxy for thermal regime shifts. - Local factors with high spatial variability (e.g., snow density, vegetation, soil thermal properties, drainage, taliks, detailed subsurface stratigraphy) were not fully incorporated; absence of such features in both training and projection datasets may limit local-scale accuracy. - The framework uses classic LSTM; while effective, newer sequence models might improve performance subject to further evaluation.