Spatial downscaling

[ethanplunkett]

BirdFlow models are constrained to relatively coarse cell sizes, and the cost of reducing the cell size scales roughly O(n^4) with the inverse of the cell dimension; halving the cell dimension results in 16 times as many parameters in the model.

The spatial resolution of the output of both predict() and route() can be improved by spatial downscaling the results. Use the S&T high resolution abundance for the timestep to assign the probability from a coarse model cell to smaller constituent cells. In the case of route() the sampled end location would then be drawn from the higher resolution raster.

This could be implemented with two functions:

1. downscale_abundance() convert low resolution abundance to high resolution by assuming birds within each coarse cell are distributed as predicted for that time in the eBird status and trends data.
2. downscale_route() Resample each location by assigning probability to nearby locations in accordance with eBird status and trends.

These approaches would be especially helpful when habitat is concentrated, such as in wetlands or coastal areas. Instead of spreading the abundance or movement end points throughout the 100 km cell we'd place it in the portion of the cell the birds are likely to occupy.