Machine learning and process-based modeling of spatiotemporal changes in active layer thickness across Alaska

Area of Science:

• Earth and Environmental Sciences
• Climate Science
• Geophysics

Background:

• Permafrost degradation in the Arctic poses significant risks to infrastructure and ecosystems due to accelerating warming.
• Active Layer Thickness (ALT) is a critical indicator of permafrost thaw and its associated impacts.

Purpose of the Study:

• To investigate the spatiotemporal dynamics of ALT across Alaska under projected climate change scenarios.
• To compare the performance of a physically based model (Stefan) with machine learning (ML) techniques for ALT prediction.
• To assess the sensitivity of ALT to future weather conditions using Coupled Model Intercomparison Project Phase 6 (CMIP6) projections.

Main Methods:

• Integrated field observations, environmental datasets, the Stefan model, and random forest (RF) ML techniques.
• Utilized CMIP6 weather projections under Shared Socioeconomic Pathways (SSP 2-4.5 and SSP 5-8.5) for future climate scenarios.
• Analyzed variable importance (mean annual temperature, slope angle, sediment transport index, stream power index) for ALT prediction.

Main Results:

• The RF model showed better performance on training data (R² = 0.84) but lower generalizability on test data (R² = 0.24) compared to the Stefan model (R² = 0.53 training, R² = 0.54 testing).
• Mean annual temperature and slope angle were the most significant predictors of ALT.
• Projected ALT increases by 2100 are 3.3 cm (SSP 2-4.5) and 5.9 cm (SSP 5-8.5) for the ML model, versus 13 cm (SSP 2-4.5) and 28 cm (SSP 5-8.5) for the Stefan model, with greatest increases in northern Alaska.

Conclusions:

• Hybrid modeling approaches combining physical and ML techniques offer valuable insights into complex ALT dynamics.
• ML models suggest more conservative ALT increases compared to the Stefan model, highlighting the importance of model selection for Arctic adaptation.
• Accelerated permafrost thaw necessitates improved predictive capabilities to mitigate risks to Arctic infrastructure, ecosystems, and carbon release.