Lakes represent as much as similar to 25% of the total land surface area in lowland permafrost regions. Though decreasing lake area has become a widespread phenomenon in permafrost regions, our ability to forecast future patterns of lake drainage spanning gradients of space and time remain limited. Here, we modeled the drivers of gradual (steady declining lake area) and catastrophic (temporally abrupt decrease in lake area) lake drainage using 45 years of Landsat observations (i.e. 1975-2019) across 32 690 lakes spanning climate and environmental gradients across northern Alaska. We mapped lake area using supervised support vector machine classifiers and object based image analyses using five-year Landsat image composites spanning 388 968 km(2). Drivers of lake drainage were determined with boosted regression tree models, using both static (e.g. lake morphology, proximity to drainage gradient) and dynamic predictor variables (e.g. temperature, precipitation, wildfire). Over the past 45 years, gradual drainage decreased lake area between 10% and 16%, but rates varied over time as the 1990s recorded the highest rates of gradual lake area losses associated with warm periods. Interestingly, the number of catastrophically drained lakes progressively decreased at a rate of similar to 37% decade(-1) from 1975-1979 (102-273 lakes draining year(-1)) to 2010-2014 (3-8 lakes draining year(-1)). However this 40 year negative trend was reversed during the most recent time-period (2015-2019), with observations of catastrophic drainage among the highest on record (i.e. 100-250 lakes draining year(-1)), the majority of which occurred in northwestern Alaska. Gradual drainage processes were driven by lake morphology, summer air and lake temperature, snow cover, active layer depth, and the thermokarst lake settlement index (R (2) (adj) = 0.42, CV = 0.35, p < 0.0001), whereas, catastrophic drainage was driven by the thawing season length, total precipitation, permafrost thickness, and lake temperature (R (2) (adj) = 0.75, CV = 0.67, p < 0.0001). Models forecast a continued decline in lake area across northern Alaska by 15%-21% by 2050. However these estimates are conservative, as the anticipated amplitude of future climate change were well-beyond historical variability and thus insufficient to forecast abrupt 'catastrophic' drainage processes. Results highlight the urgency to understand the potential ecological responses and feedbacks linked with ongoing Arctic landscape reorganization.

Anthropogenic climate change has been linked to the degradation of permafrost across northern ecosystems, with notable implications for regional to global carbon dynamics. However, our understanding of the spatial distribution, temporal trends, and seasonal timing of episodic landscape deformation events triggered by permafrost degradation is hampered by the limited spatial and temporal coverage of high-resolution optical, RADAR, LIDAR, and hyperspectral remote sensing products. Here we present an automated approach for detecting permafrost degradation (thermoerosion), using meso-scale high-frequency remote sensing products (i.e., Landsat image archive). This approach was developed, tested, and applied in the ice-rich lowlands of the Noatak National Preserve (NOAT; 12,369 km(2)) in northwestern Alaska. We identified thermoerosion (TE) by capturing the spectral signal associated with episodic sediment plumes in adjacent water bodies following TE. We characterized and extracted this episodic turbidity signal within lakes during the snow-free period (June 15-October 1) for 1986-2016 (continuous data limited to 1999-2016), using the cloud-based geospatial parallel processing platform, Google Earth Engine (TM). Thermoerosional detection accuracy was calculated using seven consecutive years of sub-meter high-resolution imagery (2009-2015) covering 798 (similar to 33%) of the 2456 lakes in the NOAT lowlands. Our automated TE detection algorithm had an overall accuracy and kappa coefficient of 86% and 0.47 +/- 0.043, indicating that episodic sediment pulses had a moderate agreement with landscape deformation associated with permafrost degradation. We estimate that lake shoreline erosion, thaw slumps, catastrophic lake drainage, and gully formation accounted for 62, 23, 13, and 2%, respectively, of active TE across the NOAT lowlands. TE was identified in similar to 5% of all lakes annually in the lowlands between 1999 and 2016, with a wide range of inter-annual variation (ranging from 0.2% in 2001 to 22% in 2004). Inter-annual variability in TE occurrence and spatial patterns of TE probability were correlated with annual snow cover duration and snow persistence, respectively, suggesting that earlier snowmelt accelerates permafrost degradation (e.g. TE) in this region. This work improves our ability to detect and attribute change in permafrost degradation across space and time.