Model experiment figures

[limnoliver]

In both the lake and stream modeling papers, the PGDL models were challenged through a number of experiments.

One experiment was to artificially reduce the number of training observations for the different model types, because one critique of machine learning is that though it performs well under data-rich conditions, can do very poorly in data-poor conditions. How does the hybrid model hold up?

A different experiment was to test the model in unseen conditions, as another critique of machine learning models is that while they can perform well under circumstances the model had "seen" during training, under new scenarios can do very poorly. The models were trained on only cold season (fall, winter, spring) data, but tested in summer.

We would like to get line (for the data sparsity test) and bar (for the seasonal test) plots for these experiments that look consistent across the lake and stream results.

• Start with the stream figures. You'll be extracting data from these tables, so the first thing you'll need to do is extract these data into a csv so they can be read into R.

• The data sparsity line plot should look vaguely like this figure, with the 1, 10, 100% on the x, model metric on the y, and each model getting their own line.

• Try the seasonal test as a bar plot. I'm not sure what will look best here. Probably color = model or color = season, with whatever is not used for color as the primary grouping variable for the bars (e.g., if color = season, bars should be grouped by model). Play around with different options, including whether the metric values should go on the x or y.

• Once you're done with these, there may be an additional task of recreating the lakes figures to look similar. Stay tuned!