It's been a fun and productive summer!

Broadly speaking, my goal for this summer was to explore the application of Machine Learning in modeling hydrological systems. Specifically, we implemented Random Forests to make short-term stream-flow forecast in a mixed snow/precipitation dominated watershed. Streamflow is hydroclimatic variable that is essential in water resource management and flood warnings.

In previous studies where ML are used to make streamflow prediction, precipitation was the only major driver. The highly heterogeneous in climatic and topographic settings of the Pacific Northwest Watersheds where both precipitation and snowmelt play important roles provide both a challenge and an opportunity to access the capability of a ML method to make streamflow forecast, thus the objective of the study.

Figure 1. Pacific Northwest Hydrological Unit

Background

Pacific Northwest is one of the 21 major national hydrological units designated by US Geological Survey. Most precipitation in this region occurs in the winter and the summer tends to be arid, especially the eastern side of the Cascades that runs north-south along the Washington-Oregon coast. A considerable amount of research have been dedicated to examine and document the long-term trends between the runoff generated by both precipitation and snowmelt using in-situ observations and model simulation.

Occupying a large area, PNW can be further divided into 12 sub-regions (Table 1). There are much variations in the climatic conditions among these sub-regions. To illustrate, the monthly average of streamflow, daily precipitation, and snow-water-equivalent during the period 2009-2018 of HUC 17-01, 17-03, and 17-09 are plotted (Fig 2).

What I accomplished.

Spatial variability in model performance

Seasonal variability in model performance

I was able to gather data, got the model running, and produced some diagnostic results. Although there were some hiccups at the beginning (lack of access to certain data, VPN not working, uncertainty about the direction of the project, etc.), I'm a little relieved to have spent the summer productively. The three datasets that Tyer and I worked on this summer (USGS streamflow, SNOTEL snow water equivalent, and PRISM gridded precipitation) open up so many potential projects in the future. I'm also a new Linux user and I'm much more comfortable with the operating system now.

I'll spend the next few weeks digging into the result and hopefully we can turn it into a paper for hydrology/meterological journal.