For ease of reference, here are the relevant paragraphs of that recent CV update blog post which described the changes in the geomodel. Warning: These are written in serious modeling geek-speak.
"When we update the computer vision model, we also update the geomodel. With this release, we’ve updated our geomodel using a new training approach called SINR (spatially implicit neural representation). Our previous geomodels predicted species distributions based on spatial grid cells. All iNat observations were aggregated into multi label presence sets per cell, then a model was trained on these aggregations using multi label binary cross entropy over species presence in the cell. It produced sharp geo priors which make it useful in downstream computer vision tasks, and is straightforward to develop and debug.
In contrast, a SINR model learns directly from individual observations and carefully constructed pseudo absences, and uses negative sampling loss to distinguish where species are likely or unlikely to occur. It provides better generalization and avoids discretization artifacts, aligns with the work of our research collaborators, and is easier to adapt to interesting new directions for producing pseudo absences. It’s empirically stronger on mapping tasks, and we believe that as we keep working on it, it will catch up on the geo prior / CV task."