Assessing potential of sparse‐input reanalyses for centennial‐scale land surface air temperature homogenisation

Abstract

AbstractObservations from the historical meteorological observing network contain many artefacts of non‐climatic origin which must be accounted for prior to using these data in climate applications. State‐of‐the‐art homogenisation approaches use various flavours of pairwise comparison between a target station and candidate neighbour station series. Such approaches require an adequate number of neighbours of sufficient quality and comparability – a condition that is met for most station series since the mid‐20th Century. However, pairwise approaches have challenges where suitable neighbouring stations are sparse, as remains the case in vast regions of the globe and is common almost everywhere prior to the early 20th Century. Modern sparse‐input centennial reanalysis products continue to improve and offer a potential alternative to pairwise comparison, particularly where and when observations are sparse. They do not directly ingest or use land‐based surface temperature observations, so they are a formally independent estimate. This may be particularly helpful in cases where structurally similar changes exist across broad networks, which challenges current techniques in the absence of metadata. They also potentially offer a valuable methodologically distinct method, which would help explore structural uncertainty in homogenisation techniques. The present study compares the potential of spatially‐interpolated sparse‐input reanalysis products to neighbour‐based approaches to perform homogenisation of global monthly land surface air temperature records back to 1850 based upon the statistical properties of station‐minus‐reanalysis and station‐minus‐neighbour series. This shows that neighbour‐based approaches likely remain preferable in data dense regions and epochs. However, the most recent reanalysis product, NOAA‐CIRES‐DOE 20CRv3, is potentially preferable in cases where insufficient neighbours are available. This may in particular affect long‐term global average estimates where a small number of long‐term stations in data sparse regions will make substantial contributions to global estimates and may contain missed data artefacts in present homogenisation approaches.