Spatially adaptive neighboring information improves constraints on local temperature projections

Adaptation needs local temperature projections, yet their uncertainty remains large. Historical temperature differences between observations and the model-ensemble mean contain unforced variability, substantially limiting constraints on future local warming. Whether neighboring records strengthen these constraints, and at what spatial scale, remains a key unresolved question. We show that distance-weighted averaging of observation–model differences within a fixed radius improves skill over using only the target-grid record. This improvement arises because nearby locations share forced warming signals, allowing averaging to reduce noise while retaining information about future warming at the target. We therefore develop an adaptive observational constraint that selects the neighborhood radius and weights neighboring information according to its noise and relevance to target warming. It improves skill over the target-grid constraint in 33 of 34 SSP5-8.5 models at a 50-year lead and in both lower-forcing SSPs. Applied to HadCRUT5, it lowers projected late-century local warming under SSP5-8.5 and narrows spread.