Aggregating Covariates in Spatial Downscaling

spatial statistics
downscaling
covariates
How covariates get aggregated to coarse observation footprints, and why the same change-of-support operator handles both covariates and the latent field
Published

March 22, 2026

## The problem

Downscaling is hard because going from coarse to fine is underdetermined. But aggregating a covariate from fine to coarse is well-posed — it’s just a weighted average. This asymmetry is worth understanding.

How covariate aggregation works

[The key idea: a high-resolution covariate like landcover can be averaged down to the observation footprint straightforwardly, using weighted averages over the footprint geometry]

The A matrix connection

[The elegant result: the same change-of-support operator A from the residual discretization provides a consistency-preserving approximation to covariate aggregation. So you don’t need a separate aggregation scheme — the geometry is already encoded in A]

\[X^{(o)} \approx A X^{(\ell)}\]

## Why this matters practically

[You can exploit high-resolution covariates without paying the computational cost of a fine latent grid. The two discretizations — fine for covariates, coarser for the residual — can be handled separately and cleanly]

In SpatialBasis

[How to pass covariates, what resolution they should be at]

Example

[Code example with a real covariate]