I am new to drivendata and noticed that so far, there are very few discussions on this forum about the different approaches to solve the problem. I am a newcomer in this field and my interest is mostly to discover it as well as trying ML time-series approaches.
I found this very nice review on SWE estimation Toward Snow Cover Estimation in Mountainous Areas Using Modern Data Assimilation Methods: A Review.
The first approach, used in most countries for SWE, is to create a physical model of the snowpack, and use hourly estimates of several parameters (mainly temperature, humidity, wind and albedo) to update these models.
In some countries such as Switzerland, many in-situ measurements are available and are used to do data assimilation: the SWE estimates of the model are corrected using the different observations. This could be done on this challenge using the SNOTEL data.
An alternative approach is to use satellite data. The optical sensors (MODIS, LANDSAT and Sentinel-2) are mostly useful to get the snow cover (is there snow or not ?) and the albedo. Snow cover is available in HRRR data (SNOWC variable here) but I have no ideas how they computed it. The Sentinel-1 data look much more promising, as its microwave backscatter signal is directly correlated with SWE (see Lievens. et al).
What approaches look the more promising for you ? Are you working only using HRRR data ? Sentinel-1 data ? SNOTEL data ? A mix of the three ? How much ML are you using in your approaches ? Would be happy to learn from your feedback !