Decoding Spatiotemporal Methane Emissions in Permafrost Catchments: A Machine Learning Approach

Methane (CH₄) emissions from permafrost catchments represent a critical component of climate feedback mechanisms. Therefore, understanding spatiotemporal dynamics and drivers of CH₄ emissions from rapidly changing permafrost regions is critical for improving our understanding of these changes. However, flux calculations must rely on methods which precisely and flexibly account for non-linear gas concentration increases when using a floating chamber system. Exponential concentration increase, convex increases, and/or step changes from ebullition occur frequently in CH₄ concentrations collected using floating chambers. This study introduces a flexible alternative to the available methods for isolating non-linear concentration increases by using the results of gradient boosting models and general additive models to calculate flux via an interactive algorithm. We used the algorithm to calculate CH₄ fluxes for 707 floating chamber measurements collected from a surface water in a permafrost catchment between May and August and over two field seasons on Disko Island, Greenland. Resulting flux calculations were visualized as heatmaps overlaid on an orthomosaic of the catchment area, revealing significant spatial and temporal patterns in CH₄ emission hotspots. Approximately 94% of CH₄ emissions were attributed to diffusive processes while the remaining 6% were attributed to processes resulting from ebullition. Because ebullitive events were statistically unpredictable, we report here on diffusive CH₄ emissions, which had a median of 0.0002 mg/m²/s^-1 and showed seasonal variability, ranging between -0.0001 and 0.02 mg/m²/s^-1, with highest emissions occurring during the thaw period and in the height of growing season. The highest uptake levels were outliers detected atop snow in 2024, but overall, there was not significant uptake across surface water. Streams connected to the lake emitted significantly higher rates of CH₄ throughout the season as compared to the surface of the lake. Gradient boosting machine results suggest emission hotspots were partially dependent on shifting environmental conditions where fluxes during the thaw season were controlled by variability in rainfall, wind direction, increasing air and soil temperatures, and high soil moisture content in the active layer (i.e., increasing surface water levels). Climatological and hydrogeological controls progressively gave way to biogeochemical controls as the system began to warm, oxidize, and utilize the 24-hour Arctic sunlight and the accumulated dissolved organic matter from the thaw period. Overall, this study provides insight into the seasonal dynamics shaping CH₄ emissions in a dynamic permafrost catchment.