Seasonal biases in glacier surface flow velocities measured from optical remote sensing images

Abstract. Sub-annual glacier surface velocity time series are receiving increasing attention as they provide critical information on subglacial hydrology, glacier and ice-sheet instabilities, and responses to external forcing. Regional-to-global scale datasets of glacier surface velocities are based on optical and/or SAR remote sensing images. They are now available worldwide in open access or on demand. In the meantime, systematic seasonal errors (i.e. biases) have been reported in time series derived from optical imagery on various landforms, such as landslides or dunes. The extent to which such biases affect glacier velocity maps remains poorly investigated. They may lead to misinterpretations of sub-annual velocity variations, for instance by amplifying or attenuating real seasonal signals, and in the worst case, by producing artificial seasonal signals. Here, we propose characterizing the amplitude and spatial distribution of seasonal biases identified in velocity maps derived from Sentinel-2 and Landsat 8/9 images. First, we compare GNSS velocity data measured on Argentière Glacier (French Alps) with three different glacier surface flow velocity datasets derived from remote sensing images and post-processed using different state-of-the-art methodologies. We highlight a systematic overestimation of the seasonal amplitude among the different datasets, regardless of the post-processing strategies employed. Then, we assess this seasonal bias at the scale of the Mont-Blanc Massif by analyzing velocity estimated over static areas. Clear seasonal biases are observed on the north-facing slopes, particularly on the northwestern aspect. The amplitude of this bias differs between the existing glacier surface velocity products, depending on the satellite images used and filters applied on the raw data. Across the Mont-Blanc Massif, the median value of this bias ranges from 7 to 80 m.yr−1 on north-facing slopes, for high to low level of filtering respectively. This highlights the importance of filtering the raw velocity data based on a wide range of filters, including spatial and directional filters, to reduce the amplitude of the seasonal bias. However, the measured velocities still reflect shadow tracking in areas affected by shadow casting. The amplitude of the seasonal bias is positively correlated with the number of days under shadows, and negatively correlated with variations in illumination. Therefore, to mitigate this seasonal bias, we propose a strategy to mask out areas affected by shadow casting. Finally, we estimate the amplitude of the seasonal bias in the remaining glacier covered areas based on an analysis over static areas. In the end, we show evidences of seasonal biases in other places of the world, such as Yukon, Canada and Western Greenland. The code is published as a Python package called SeBVel.