Snotel

DrivenData is a website that hosts machine learning competitions, often involving some component of Earth Science or hydrology. In 2022, they hosted a competition sponsored by the US Bureau of Reclamation to predict snow quantity across the Western US. Although I didn’t have much experience with machine learning at the time, I decided to participate in the competition and was pretty surprised to rank 62 out of 1000 using a simple approach. See these three blog posts for more details: ...

A few months ago I found out my local ski area, Sandia Peak, preemptively chose not to open for the upcoming ski season (2022/2023). It isn’t that unusual for Sandia Peak to stay closed for the season, and they were also closed last season. However, I had thought the decision to stay closed is typically made around January, after the snow pack begins to form. At that point, if the early season snowfall is too low, the season isn’t long enough offset operating costs and it isn’t worth opening. Why was this year different? ...

Overview Last winter I tried my hand at competing in a machine learning competition to predict snow water equivalent (SWE) across the Western United States. I learned a lot and created a two part blog series to document both the competition and my approach: Machine Learning for Snow Hydrology - A Competition Machine Learning for Snow Hydrology - Methods I competed in the preliminary phase of the competition that didn’t include any prizes. The second phase involved predicting SWE in real time and included big prizes totaling $500,000. The competition ended in early summer, but the winners were just recently announced on the DrivenData Blog: ...

This is the second part of my two part series on a machine learning competition to predict snow water equivalent (SWE). In Part 1, I describe the competition, as well as, my process for coming up with an approach for making SWE predictions at 9,067 locations across the Western US. That approach, sometimes called the “hypsometric” method (Fassnacht et al., 2003, see Part 1 for an overview of the method), is one of the easiest I could find, and it therefore seemed doable given personal time constraints. My expectations were low - I just wanted to see how a simple approach compared to others in the competition. To my surprise, out of about 1000 predictions submitted to the competition, my predictions ranked 62. Here I describe how I computed the SWE predictions and assess the results. ...

Part 1: Competition Overview Late last December I ran across a machine learning competition hosted by Driven Data. The goal of the competition is to predict snow water equivalent at high spatial resolution across the western US. I had never before thought of participating in a machine learning competition, although I had heard of the idea via another platform, Kaggle. However, a machine learning competition involving snow is more up my alley, as I have both professional and personal experience with snow science. Furthermore, I had been wanting to enhance my familiarity with machine learning techniques. I decided to give it a shot. ...