Evaluating PlanetScope UDM 2.1 for High-Resolution Snow Cover Monitoring in the Arctic

By Geoffrey de Graeve

Abstract
This study evaluates the performance of PlanetScope’s Usable Data Mask 2.1 (UDM 2.1) for monitoring snow cover dynamics and surface phenology changes in Arctic regions. By comparing ground-based snow cover observations to satellite derived UDM 2.1 predictions, the accuracy and limitations of using UDM 2.1 for start of season (SoS) and end of season (EoS) were investigated. Despite a high revisit time (daily) and spatial resolution (3 x 3 m), results show that image resampling poses challenges for small scale snow cover change detection, as the UDM 2.1 pipeline reduces the spatial resolution by a factor 7.5. Temporal gaps in image availability, particularly close to the EoS, further limit its use for EoS detection, making it impossible using solely data provided by Planet. Uncertainty in predictions at class boundaries introduces additional challenges, causing some locations to be labeled as clear 2 months before the location was first observed to be snow free in ground observations. Nonetheless, predictions for first snow free day (FSFD) achieved a mean absolute error (MAE) of 7.79 days and a mean error (ME) of 7.60 days using confidence thresholding at 50%. Predictions for first snow free week (FSFW) performed at a MAE of 5.39 days and a ME of 4.5 days under 40% “Clear” and 80% “Snow” prediction confidence thresholds. Although still outperformed by earlier research, it still shows potential of monitoring snow cover at a broader scale. Alternative NDVI-based methods were explored and resulted in a MAE of 4.45 days and RMSE of 4.96 days, slightly outperforming the UDM 2.1. Using thresholding for SoS predictions in Adventdalen resulted in an SoS of roughly 7 days earlier than yearly average for 2000-2019 using both UDM 2.1 and NDVI. While PlanetScope UDM 2.1 offers valuable high-frequency imagery, its current processing pipeline and classification strategy impose critical limitations for snow cover and seasonal phenology change detection at fine spatial and temporal scales. Incorporating NDVI and recent developments in Machine Learning could enhance its reliability in monitoring of the Arctic.