Climate Field Reconstruction Under Station Sparsity and Geographic Shift

Uneven station coverage limits regional analysis of temperature, precipitation, and climate extremes. We test whether compact neural models improve same-day reconstruction from European stations and remain reliable under station sparsity and geographic shift. Using ECA&D observations and ERA5 reference fields, we compare a latent-grid transformer and ConvCNP with tuned local kriging across eight training seeds. The transformer reduces in-domain temperature RMSE from $2.21$ to $1.16,^\circ\mathrm{C}$, while precipitation RMSE is smaller, from $3.80$ to $3.45$ mm/day. Under eastward geographic transfer, kriging remains near $2.24,^\circ\mathrm{C}$, whereas neural errors rise to $4.12$--$4.73,^\circ\mathrm{C}$. Conformal recalibration restores coverage only through wider intervals. Neural models can improve reconstruction in familiar regions, but geographic shift must be evaluated before reconstructed fields support climate analysis.