Thanks for the questions - happy to clarify.

• Using latitude/longitude to extract TerraClimate (or other provided environmental raster) values is allowed. The restriction is on using latitude/longitude themselves (or features derived from them) as model inputs. In other words: ✅ Using coordinates to sample the provided climate rasters ❌ Using coordinates, distances, grid cells, clusters, nearest neighbours, or other spatial encodings as predictive features
• Derived ecological/climate features computed purely from the extracted climate variables are allowed. For example: climate indices seasonal aggregations water balance variables interactions between TerraClimate variables are all fine, provided they are derived only from the allowed environmental data.
• Pseudo-labeling is permitted only if it uses predictions generated from compliant models and the provided competition data. However, participants should avoid approaches that indirectly leak information from the test set or reconstruct labels using external spatial information.
• Regarding timestamps and temporal leakage: Participants should avoid using information that would not realistically have been available at the time of observation. In practice, using climate summaries that overlap the observation period is acceptable if they come from the provided TerraClimate data, but care should be taken not to introduce future information.

Hope that helps!