How can we use a weekly snow cover time series to improve glacier monitoring and modeling in mountainous areas and predict water storage in foothills and lowlands?
Seasonal snow and ice melt strongly influence glacier mass balance, yet sparse sub‐annual observations limit our understanding of seasonal dynamics. Here we construct and analyze weekly snow cover time series for 200 glaciers across western North America from 2013 to 2023 using an automated image processing pipeline. Snow cover varied widely across the region: snow minima timing varied with latitude — from ∼August from 62 to 64°N to ∼October from 48 to 50°N—and accumulation area ratios ranged from near‐ zero to 0.92 (median of 0.52). A comparison of snowlines from observations and the PyGEM glacier mass balance model revealed seasonally evolving but spatially consistent biases in modeled snowlines: observed snowlines rose earlier, but at a slower rate throughout the melt season, than modeled snowlines. Beyond capturing glacier state, snowline observations efficiently provide sub‐seasonal mass balance constraints and empirically represent unresolved processes like snow redistribution, refining model gradients and improving projections. Plain Language Summary Glacier models perform best when calibrated to match observations. Although model calibration commonly focuses on decade‐scale glacier change, new advances in satellite image processing and image coverage enables the creation of detailed observations of glacier snow cover that may be useful to further tune models. Using approximately weekly satellite images to track snow cover on 200 glaciers in western North America from 2013 to 2023, we found that the average timing and extent of minimum snow cover varied greatly between glaciers. The seasonal evolution of glacier snow cover was not accurately captured by a state‐of‐the‐art model used for global glacier mass loss projections. However, for a subset of glaciers, we show that the difference between modeled and observed snow cover can be reduced by adjusting the modeled temperature and rate of melting with each degree above freezing. Our study emphasizes the value of detailed satellite‐based snow cover time series for observing key glacier health metrics and improving glacier model predictions.Broad glacier mass loss in the 21st century is projected consistently across global glacier models, yet uncertainties remain in the exact magnitude due in part to differences in the representation of glacier‐climate interactions (Marzeion et al., 2020). Glacier massloss outside of the ice sheets has been primarily driven by changes in surface mass balance (SMB; i.e., net difference between snow accumulation and melt; Larsen et al., 2015; Menounos et al., 2019), such that projections of glacier mass change are strongly reliant on the accuracy of SMB models. These models are often calibrated using temporally limited glacier surface elevation and mass balance observations, reducing their accuracy and contributing to uncertainty in long‐term projections of glacier mass balance (Hock et al., 2019; Marzeion et al., 2020). Satellite imagery has revolutionized our ability to monitor critical glacier mass balance metrics from space, providing valuable constraints for SMB models. For mid‐latitude glaciers, the seasonal snowline at the end of the melt season is commonly used to estimate the equilibrium line altitude (ELA; Østrem, 1975; Rabatel et al., 2005), the altitude where annual net SMB is zero (Cuffey & Paterson, 2010). Similarly, the accumulation area ratio (AAR) indicates the portion of the glacier with positive SMB, and is calculated as the snow‐covered glacier area divided by the total glacier area, typically at the end of the melt season. Both metrics have been mapped using satellite imagery on regional scales (e.g., Larocca et al., 2024; Liu et al., 2021; McFadden et al., 2011; Zeller et al., 2025). However, observations are often limited by cloud cover, recent snowfall (Rabatel et al., 2005), and sparse repeat intervals for individual satellites, confounding the identification of annual minimum snow cover. What's more, spatiotemporal variability in the seasonal timing of snow cover minima has not been interrogated on regional to global scales. Given the recorded and projected shifts in the timing of snow accumulation and melt (Barnett et al., 2005; Littell et al., 2018; McCrary et al., 2022; Mote et al., 2018; Musselman et al., 2021; Rhoades et al., 2022; Siirila‐Woodburn et al., 2021), it is increasingly important for observation‐based glacier analyses to account for changing snow minima timings when designing sampling strategies and calibrating glacier mass balance models. Here we apply a machine learning‐based image classification pipeline (Aberle et al., 2025) capable of distinguishing glacier ice and firn from seasonal snow to create approximately weekly time series of glacier snow cover for 200 glaciers in western North America for 2013–2023 (Section 2.1). The pipeline was designed to process imagery from both the Landsat 8/9 and Sentinel‐2 satellite constellations, helping to mitigate limitations imposed by the repeat intervals of individual sensors. We focus on western North America because it has the highest regional mass loss rate outside the ice sheets (Gardner et al., 2013; Hugonnet et al., 2021; Zemp et al., 2019), it is projected to continue to be a primary contributor to global glacier mass loss throughout the 21st century (Marzeion et al., 2020; Rounce et al., 2023), and it includes diverse climate zones. To account for the region's diverse climate zones, we cluster the sampled glaciers using decade‐long climate reanalysis data (Section 2.2), allowing for clearer interpretation of snow cover trends across temperature and precipitation regimes. We apply the pipeline to Landsat 8/9 and Sentinel‐2 images from 2013 to 2023, creating a robust data set to evaluate spatio‐temporal variations in glacier snow cover minima and assess seasonal to annual biases in hindcasting simulations for prognostic glacier mass loss models (Section 2.3).2. Methods and Materials 2.1. Remotely‐Sensed Glacier Snow Cover Time Series Snow cover time series were automatically constructed at each glacier using the glacier-snow-covermapping pipeline (Aberle et al., 2024). The pipeline was developed by training and testing several supervised machine learning algorithmsto classify snow, shadowed snow, ice/firn, rock/debris, and water in Landsat 8/ 9 and Sentinel‐2 imagery. For this study, all shadowed snow pixels were set to snow. The image classifiers were trained on ∼32,000 manually classified points spanning the melt season at four glaciers using K‐fold cross‐ validation. The best classifiers for each image product were the Nearest Neighbors and Support Vector Machine classifiers, which achieved classification accuracies of 92%–97% when validated on ∼5,000 manually classified points at two glaciers excluded from the training set (Aberle et al., 2025). We selected 200 glaciers in western North America for our study (Figure 1a) to take into account the distributions of glacier geographic locations, elevations, aspects, slopes, and areas in Randolph Glacier Inventory (RGI) version 6 regions 1 and 2, excluding the Brooks Range (RGI Consortium, 2017). Focusing on this subset enabled exports of the classified image time series for a benchmark data set for further analysis of snow distribution patterns, which was not possible region‐wide due to Google Earth Engine's export limitations. We excluded the largest glaciers (areas >5,000 km2 ), which make up < 0.1% of all glaciers by number in the region, due to limited cloud‐free imagery covering the full glacier area. Glaciers smaller than 0.1 km2 were also excluded, which comprise about 20% of the regional glacier population by number (