sdtaylor/bbs-sdm

bbs-sdm's Issues

paper to do list

Things to do before finalizing this project for the scale paper

• lighten up the restrictions on including sites. now as long as a grid cell has the minimum # of sites it can be included, instead of doing things on a site by site basis

• better documentation throughout so others can read it

• potentially use raster::aggregate() for scaling up prism data instead of doing it manually

• switch to random forest instead of gbm.

• re-run completely from scratch, including download data, and processing

• make sure BBS models are not aggregating observations

add support for multiple nodes on hipergator

To scale past 32 cores it would help a lot to be able to run the script distributed across many nodes. If I end up running the full analysis more than once a week this would be a requirement.

Test niche conservation

ENMTools: a toolbox for comparative studies of environmental niche models

ENVIRONMENTAL NICHE EQUIVALENCY VERSUS CONSERVATISM: QUANTITATIVE APPROACHES TO NICHE EVOLUTION <- seems pretty popular

make results be written to postgress instead of csv file

consider other models

Models currently configured are randomForest, GLM, CTA, and GBM. Currently only using GBM as it's the best out of those 4. Other models to consider:

• GAM
• MaxEnt (presence only)
• OneClassSVM (presence only)
• MaxLike (presence only)

These state of the art models should theoretically perform better than any of the above.

• MistNet
• Gaussian process

Add simulation modeling

make a simulation model of species responding (or not) to various environmental forcing

Likely it's own project, and opens up a big can of worms.

improve comments

less redundant comments. more comments in function descriptions.

decide on performance assesment

Many different ways to quantify the usefulness of an SDM.

• Rapacciuolo et al. 2012 compared compared only plots that changed between two time periods. He used the correct classification rate (same as precision).
• Rapocciulo et al 2014 conceived Temporal Validation Plots specifically for comparing SDM's across two time periods. They use all sample points available whether they changed or not between the two periods. They are heavily weighted toward points that do change though. They are a plot that is meant to be interpreted visually for individual spp, but can be summed up in custom functions for precision, correlation, and bias.
• Looking at any single metric (auc, precicion, ccr) over time is extremely noisy due to large environmental variation. Even at the larger temporal scales. This is potentially smoothed out by gradually offsetting the training years and averaging things out. see #4
• Looking at the observed changes of sites with high modeled probabilities of change between two time periods (extreme deltas) seems to gives a very intuitive feel for SDM performance over time.

add utils file

For things like bioclim processing that doesn't change that much.

add spatial scale

See what increase in spatial scale does, as opposed to single points that the BBS routes represent.

Do this by drawing geographic square or rectangles across the US and having species be present/absent in them. Test the spatial scale by gradually increasing those squares.

Some initial logistic issues:

• How to compare densely sampled eastern US with the sparsely sampled west?
• Need a minimum number of routes that need to be in a square.

add yearly offsets

To account for climatic variability, add offseting to gradually shift the training time period forward in time.