search

sdtaylor / phenology_dataset_study

1 stars 1 forks source link

on when parameters predict no budburst/flowering #27

Closed sdtaylor closed 6 years ago

sdtaylor commented 6 years ago

It's possible, with all the models, to have a set of parameters which predicts an event to happen well into the winter. This is obviously unrealistic. In the model fitting routines if the differential evolution algorithm tries out parameters which predict an event any time past around ~Oct 15, the doy estimate is instead set to 1000, which returns a very large error and those parameters are seen as non-optimal by DE.

This works very well until the fitted model is used on out of sample data. For example when I take NPN built models and try to predict Jornada data I sometimes get doy 1000, which is basically saying the model predicts the event will not happen that year.

How to deal with this? I don't know. In all the phenology literature virtually everyone uses the same setup (usually simulated annealing to minimize RMSE), but only one (Olsson & Jonsson 2014) mention this issue (but not how they deal with resulting predictions).

sdtaylor commented 6 years ago

I think I will just drop any predictions that end up being DOY 1000. Luckily this is only a very very small subset. Of the instances where it happens, only some of the 250 bootstrapped predictions for any 1 observation have this issue. So dropping these non-predictions should not affect much.

sdtaylor commented 6 years ago

image

Here are the instances where this happens. Interestingly it's mostly the alternating model. Larrea is a precip driven phenology so it's no surprise that has issues. The HJ andrews species could likely be ones driven mostly by snowmelt, especially Trillium (a small forb), but I'm not sure sure about Rhododendrun and Vaccinium (shrubs).

sdtaylor commented 6 years ago

this is a very interesting thing to look into further, as all models I know of only predict when some phenological event will occur, not if it will occur in a given season. This obviously doesn't fit a lot of desert plants.

sdtaylor commented 6 years ago

dropping "non-predictions" https://github.com/sdtaylor/phenology_dataset_study/commit/9bf6a62eb760a829ac5bd9e05d6fa27ad284c388