Our let-it-snow (LIS) processor computes the extent of the snow cover from Sentinel-2 surface reflectance (level 2 products). However, it can also compute multi-temporal composites (level 3 products) such as the snow cover duration, snow onset and snow melt out date from a time series of snow cover maps with a cloud mask. Aurore Dupuis (CNES) has upgraded this module to handle the snow products from the Copernicus High Resolution Snow & Ice Monitoring Service (HRSI). Since the HRSI products are generated (by MAJA and LIS as well) at pan-European scale, we are now able to generate these composites over any mountain range in Europe at 20 m resolution for the past four years. In general we consider the hydrological year as the period of aggregation, i.e. from 01 September of year N to 31 August of year N+1  (read the method).

Here is an overview of the products over the Alps and a zoom near Barcelonette.

Note that the « snow onset date » is noisy and the imprint of the Sentinel-2 swath is still visible in some areas. This reflects the high sensitivity of this indicator to the observations timing. Indeed, the « snow onset date » occurs in early winter when… it’s snowing. And when it’s snowing, there are clouds in the sky hiding the land surface. The start of the snow season is therefore rarely well captured by optical sensors such as Sentinel-2 despite its revisit of 5 days.

Arthur Bayle (LECA) will use the data for an analysis of vegetation trends in the Alps in the frame of the TOP project funded by the French Research Agency. I think we can learn many things from this dataset. I already have some ideas… What would you do?

5 thoughts on “Multitemporal snow metrics from Sentinel-2

  1. Hi Simon,
    Thanks for this post! I did not know that the number of observations varied so significantly for orbital reasons. Something good to know for the users, when they pick a study area!
    I also think that we can learn a lot from the SMOD for model evaluation because it’s an ‘integrate’ over the season. But due to the limited revisit frequency/cloud coverage, I guess that the uncertainty of a pixel’s SMOD can amount to days: does your SMOD product come with an estimation of this uncertainty? That would be a plus in situations where the model performance is already pretty good 🙂

    1. Assuming that the snow/no-snow detection is perfect, the uncertainty on the SMOD is the revisit time of Sentinel-2. But this revisit actually varies spatially due to orbital parameters and cloud cover! This is also why I provide NOBS, to have an idea of the actual revisit. Hence you can estimate the uncertainty as 365/NOBS.

      1. Also note that the NOBS pattern is not visible in the SMOD and SCD maps, which is a good sign that the sampling frequency is sufficient to capture the phenomenon we look at. In the details there are many defaults though!

          1. not necessarily, but their magnitude is lower than the signal at this scale at least!

Laisser un commentaire

Votre adresse e-mail ne sera pas publiée. Les champs obligatoires sont indiqués avec *

Ce site utilise Akismet pour réduire les indésirables. En savoir plus sur comment les données de vos commentaires sont utilisées.