Go SWIFTly into that good night

I'm in Canberra this week for a research planning meeting. Each of about 12 research leaders have to say something about their progress and plans on work to support the Bureau of Meteorology. I gave a talk about my work with developing better short-term (hours ahead to days ahead) river flow forecasts. The organizers encouraged everyone to bring a prop to support their talk. I wrote this poem, "Go SWIFTly into that good night"

The rivers rose throughout the night
the users woke in such a fright
The rain came down like cats and dogs
the channels choked with fallen logs
“What should we do?” they cried aghast
“Lets hope this night is not our last”
It all begins with NWP
to say what future rain will be.
Not knowing that will leave you sunk
too bad the results always stunk.

NWP is Numerical Weather Prediction, weather models. In most cases, you can say very little about future floods without a good weather forecast. Unfortunately, these models are not good about predicting rain (as opposed to temperature).

If by now you aren’t defeated
find out if the soils’ depleted.
Simulate a little bit and
find parameters with good fit

We use hydrology models to track how dry the soils are and what effect it'll have on streamflow production. These models have parameters that need to be calibrated.

But don’t catch bad habits like
helping models breed like rabbits.
Existing structures are just fine
unless you need to say “all mine!”

Most hydrology models were developed in the 1960-1970s. Over the years, everyone and his sister has made their own minor variant of these models and gave it their own name. It's a bit cynical to call these "vanity models" but often times only the developer ends up using their own model. This didn't get nearly the laughs I was expecting, probably because many of the people in the room were guilty of this.

GR4J seems fine to me
but demonstrating skill is key.
Is there an incremental gain?
Or are the results just the same?

GR4J is a great french model that is simple and works well.

 
We must say how uncertain we are
especially in leadtimes far
If the forecasts are off kilter
I recommend a kalman filter
Or maybe try your best and let
error correction clean the mess.

A Kalman filter is useful for looking at how well your forecasts are tracking the recent observed, and if there are differences, it'll say how to fix the model states. I couldn't think of a rhyme with 4 Dimensional Variational Assimilation.

We present wrapped like in a gift
as some software we now call SWIFT
SWIFT makes forecasts to be trusted
SWIFT won’t leave you feeling busted

SWIFT is the acronym of the software we're developing.

There’s one block to this thing
that keeps us from our researching
Others sure do have it easy
The whole ordeal makes me queasy
My wife slaves away very late
Squadrons of Anzac cookies she must bake
I beg for sub-daily data
and the Bureau says “uh… see you later!”

Lots of projects use daily rainfall data, but good sub-daily (i.e. hourly) data is hard to come by. Kitty bakes cookies for me to give away at work in exchange for data, but progress has been slow.... Notice how only an american could rhyme data ("day-tor") with later ("lay-tor"), as opposed to "dah-tah" and "lah-tah".

Well, if you agree to our demands
I ask now to join your hands
Give it up and cheer like fools
For Short-term Water Information Forecasting Tools!

Believe it or not I won a bottle of wine for best prop!