Impact of observation network non-uniformity on precipitation forecast verification and skill scores

Evaluation of short-range precipitation forecasts is sensitive to the spatial non-uniformity of surface observing networks. Gridded NWP precipitation forecasts are commonly verified against AWS/ASOS rain-gauge observations using point-based binary verification. However, this method is affected by a grid-to-point representativeness mismatch and an uneven station distribution, which can bias domain-aggregated verification scores. Consequently, domain-aggregated verification scores can be disproportionately influenced by observations from densely monitored areas.
Beyond sampling biases arising from network non-uniformity, point-based binary verification is also prone to the double-penalty effect. Displacement of precipitation features can be counted simultaneously as a miss at observation sites and a false alarm nearby. Collectively, these limitations motivate verification frameworks that better reflect the spatial nature of precipitation.
Here we assess how observation-network non-uniformity and verification configuration influence reported precipitation forecast skill over the Korean Peninsula. As a practical step to reduce network-density effects, we interpolated point-based precipitation into a gridded observation field and performed area-based verification in parallel with conventional point-based verification for an annual set of short-range forecasts. The results show that reported skill is highly sensitive to the verification framework; for some precipitation thresholds, the area-based approach yields Critical Success Index (CSI) values that are approximately 50% higher than those from point-based verification. The findings highlight that verification design can materially affect the interpretation of precipitation forecast performance in non-uniform observing networks and underscore the need for spatially representative verification frameworks.