search

holoviz-topics / EarthML

Tools for working with machine learning in earth science
https://earthml.holoviz.org
BSD 3-Clause "New" or "Revised" License
94 stars 21 forks source link

WIP: ENH: add quality prediction to Carbon Flux #23

Open stsievert opened 5 years ago

stsievert commented 5 years ago

This adds a prediction from all global sites. This is still a work in progress: I need to make this a better predictor.

stsievert commented 5 years ago

I need a baseline to start of with. To do this, I used a linear regressor on every site to predict the last year at the same site. I used the same model as a baseline with the global prediction, but predicted a different (unseen) site, not the latest year.

Apparently, the correlation coefficient is the metric cared about. I run these simulations 179 times (once for each station). Here are the distribution of correlation coefficients from these trials:

screen shot 2018-10-01 at 9 23 19 am

We can see that the linear model at one site significantly outperforms the global linear model.

The summary statistics are

Statistic One site predicting last
year at same site		 All sites predicting
one held out site
Median 0.573 0.441
Mean 0.497 0.398
jbednar commented 5 years ago

Sounds promising. Still WIP?

stsievert commented 5 years ago

Still a WIP. Mostly, I know the baseline performance (what's considered "good"), and have a metric to improve upon.