Andrew Wilcox

Active layer thickness (ALT) is a key indicator of permafrost response to environmental change. However, its long-term variation across regions can be driven by multiple geospatial factors whose relative contributions remain difficult to disentangle. In this work, we propose a framework combining two processes: long-term trajectory classification and causal inference. Accordingly, we introduce the Active Layer Dynamics Index (ALDI), categorizing ALT time series into six classes of direction and rate of change: Rapid and Gradual Thickening, Rapid and Gradual Thinning, Transitional and No Trend. ALDI relies on nonparametric trend detection, reversal identification, and bootstrap uncertainty quantification. We then conduct causal inference using site fixed-effects and first-difference estimators to examine how environmental drivers vary across ALDI classes. Applied to 129 CALM network sites (1990–2024), ALDI identifies a 3.5:1 thickening-to-thinning ratio across the pan-Arctic and substantial regional heterogeneity. In terms of causality, thermal forcing dominates at thickening sites but is absent in thinning and transitional regimes, where hydrological and disturbance processes prevail. These class-dependent relationships were not detectable in pooled analyses, underscoring the need for trajectory-aware analysis in permafrost research.

Active layer thickness (ALT) is a key indicator of permafrost response to environmental change. However, its long-term variation across regions can be driven by multiple geospatial factors whose relative contributions remain difficult to disentangle. In this work, we propose a framework combining two processes: long-term trajectory classification and causal inference. Accordingly, we introduce the Active Layer Dynamics Index (ALDI), categorizing ALT time series into six classes of direction and rate of change: Rapid and Gradual Thickening, Rapid and Gradual Thinning, Transitional and No Trend. ALDI relies on nonparametric trend detection, reversal identification, and bootstrap uncertainty quantification. We then conduct causal inference using site fixed-effects and first-difference estimators to examine how environmental drivers vary across ALDI classes. Applied to 129 CALM network sites (1990–2024), ALDI identifies a 3.5:1 thickening-to-thinning ratio across the pan-Arctic and substantial regional heterogeneity. In terms of causality, thermal forcing dominates at thickening sites but is absent in thinning and transitional regimes, where hydrological and disturbance processes prevail. These class-dependent relationships were not detectable in pooled analyses, underscoring the need for trajectory-aware analysis in permafrost research.

We present the Alaska Coupled Observations of Land-atmosphere Dynamics (Alaska-COLD) dataset, comprising hourly air and soil temperature measurements from 12 sites across Alaska's north-central regions (64°–70°N). Each site measures air temperature and soil temperatures at four depths (0–0.7 m). Seven sites provide 714–726 days of continuous records (Summer 2023–2025) capturing multiple freeze-thaw (FT) cycles. We classify FT phenology into thawing, thawed, freezing, and frozen phases using surface temperature observations, enabling phase-specific analysis of thermal dynamics. We further derive FT metrics, including n-factors, degree-day totals, and damping ratios, to characterize surface energy exchange and subsurface thermal attenuation. The records reveal strong thermal damping with depth: surface-soil diurnal amplitudes range from 0.7–4.3°C, while the deepest sensors show only 0.004–0.047°C variation across the mean daily cycle. At the annual scale, ground-temperature ranges decrease from 11–25°C near the surface to 3–5°C at 45–75 cm depth, compared with air temperature ranges of 36–45°C. Thawed periods range from 39 to 92 days annually, and air-ground coupling vary strongly by FT phase, with R2 increasing from 0.02 during thawing to 0.69 during thawed conditions. A concise comparison with colocated ERA5-Land estimates shows stronger agreement for air temperature than for soil temperatures, highlighting the value of Alaska-COLD for resolving site-level ground thermal dynamics in permafrost environments. Alaska-COLD is accessible via: https://doi.org/10.5281/zenodo.17980271.

Active layer thickness (ALT) is a key indicator of permafrost response to environmental change. However, its long-term variation across regions can be driven by multiple geospatial factors whose relative contributions remain difficult to disentangle. In this work, we propose a framework combining two processes: long-term trajectory classification and causal inference. Accordingly, we introduce the Active Layer Dynamics Index (ALDI), categorizing ALT time series into six classes of direction and rate of change: Rapid and Gradual Thickening, Rapid and Gradual Thinning, Transitional and No Trend. ALDI relies on nonparametric trend detection, reversal identification, and bootstrap uncertainty quantification. We then conduct causal inference using site fixed-effects and first-difference estimators to examine how environmental drivers vary across ALDI classes. Applied to 129 CALM network sites (1990–2024), ALDI identifies a 3.5:1 thickening-to-thinning ratio across the pan-Arctic and substantial regional heterogeneity. In terms of causality, thermal forcing dominates at thickening sites but is absent in thinning and transitional regimes, where hydrological and disturbance processes prevail. These class-dependent relationships were not detectable in pooled analyses, underscoring the need for trajectory-aware analysis in permafrost research.

We present the Alaska Coupled Observations of Land-atmosphere Dynamics (Alaska-COLD) dataset, comprising hourly air and soil temperature measurements from 12 sites across Alaska's north-central regions (64°–70°N). Each site measures air temperature and soil temperatures at four depths (0–0.7 m). Seven sites provide 714–726 days of continuous records (Summer 2023–2025) capturing multiple freeze-thaw (FT) cycles. We classify FT phenology into thawing, thawed, freezing, and frozen phases using surface temperature observations, enabling phase-specific analysis of thermal dynamics. We further derive FT metrics, including n-factors, degree-day totals, and damping ratios, to characterize surface energy exchange and subsurface thermal attenuation. The records reveal strong thermal damping with depth: surface-soil diurnal amplitudes range from 0.7–4.3°C, while the deepest sensors show only 0.004–0.047°C variation across the mean daily cycle. At the annual scale, ground-temperature ranges decrease from 11–25°C near the surface to 3–5°C at 45–75 cm depth, compared with air temperature ranges of 36–45°C. Thawed periods range from 39 to 92 days annually, and air-ground coupling vary strongly by FT phase, with R2 increasing from 0.02 during thawing to 0.69 during thawed conditions. A concise comparison with colocated ERA5-Land estimates shows stronger agreement for air temperature than for soil temperatures, highlighting the value of Alaska-COLD for resolving site-level ground thermal dynamics in permafrost environments. Alaska-COLD is accessible via: https://doi.org/10.5281/zenodo.17980271.

This paper describes a high performance, distributed artificial intelligence based approach for Arctic focused multivariate geospatial feature prediction to enhance weather model accuracy and computational efficiency through distributed dataset preprocessing. The proposed approach was tested using an Alaska wide time series air temperature dataset spanning 25 years. The workflow partitions the dataset into spatio-temporally independent chunks, which are preprocessed and scored locally. Feature scores from each chunk are then aggregated using a weighted voting average, ensuring that features consistently important across multiple regions are prioritized while maintaining global consistency. This distributed preprocessing offers benefits such as reduced computation time, enhanced model accuracy, improved generalization of AI models, and a more robust representation and understanding of complex geospatial data. Eight feature selection techniques covering filter, wrapper, and embedded methods were evaluated to identify the most relevant predictors, and their computational performance compared across distributed partitions. Four AI models were selected from linear, nonlinear, time-series machine learning, and non-time-series deep learning categories to predict air temperature and trained using databases of diverse regions from Alaska. The impact of the eight different feature selection techniques on model performance was evaluated using MSE, sMAPE, KGE, MPE (%), and prediction time. Results show that distributed preprocessing not only substantially reduces preprocessing time but also produces equivalent or better predictive performance compared to serial preprocessing. By capturing localized features and applying weighted aggregation, the workflow ensures global feature consistency while preserving regional significance. This leads to improved model generalization across heterogeneous Arctic regions and more accurate representation of spatially diverse geophysical patterns. Methods such as Kendall+DT and Pearson+MI exhibit superior scalability with increasing dataset size and computational resources, confirming their effectiveness for large-scale Arctic datasets.

Accurately predicting active layer thickness (ALT) is critical for assessing permafrost stability and mitigating risks to Arctic infrastructure, ecosystems, and communities. However, large‐scale ALT prediction remains challenging due to limited in situ observations and high spatial variability in environmental drivers. This study introduces a data‐driven framework for maximum annual ALT prediction up to +5 years in the future that combines field measurements with geospatial features. First, key ALT predictors are identified using a robust multi‐stage feature selection strategy from a large set of 80 geospatial features. Second, a weighted ensemble model of CatBoost, Extremely Randomized Trees, and Bagging regressors is trained separately at each prediction horizon. Accordingly, the proposed prediction model is evaluated across 115 sites from the Circumpolar Active Layer Monitoring network. Results highlight seven essential drivers of ALT namely geographic, seasonal, and vegetation‐related features across the Arctic. In addition, model performance consistently outperforms baseline models and demonstrates strong predictive accuracy and generalization across diverse Arctic permafrost environments.