Expanding the Swiss‐Norwegian chironomid training set with Polish data

Abstract

The Swiss‐Norwegian training model established in 2011 has been used across Europe to infer past mean July air temperatures using Chironomidae subfossils. It combines reference sites from Norway and the Swiss Alps but lacks representative localities from the central European plains belt. This gap complicates reconstructions from temperate European latitudes, and thus, a Polish training set (TS) was developed. Data from 102 lakes, mainly from Pomerania and the Masurian Lakelands, as well as the Polish Plain, the Sudetes, and the Carpathian Mountains were included. After selecting and merging suitable chironomid taxa, the Swiss‐Norwegian‐Polish TS includes 357 lakes and 134 chironomid morphotypes, and it has a mean July air temperature range of 3.5 to 20.0 °C. The weighted averaging‐partial least squares (WA‐PLS) and artificial neural network (ANN) transfer function based on the extended TS performs well: R2jack = 0.91/0.95 and root mean squared error of prediction (RMSEP) = 1.39 °C/1.34 °C, respectively. The transfer function applied to the Żabieniec Lake sequence (central Poland) reveals higher summer temperatures (2.5–4.0 °C) for the Lateglacial and for the Holocene (1.0 °C) compared to the Swiss‐Norwegian model. The reconstruction based on the Swiss‐Norwegian TS shows an increasing trend from the Bølling‐Allerød Interstadial to the Meghalayan period, whereas the new model reconstruction shows no increasing trend when using WA‐PLS and a slightly decreasing trend when using ANN. Summer temperatures in the Lateglacial in Żabieniec inferred using the Swiss‐Norwegian TS are consistent with other temperature reconstructions for 50–55°N in Europe, while those based on the Swiss‐Norwegian‐Polish TS (SNP TS) show closer affinity to more western sites than the eastern European ones. Despite being uncommon, the application of ANN can be a good equivalent of the WA‐PLS technique especially in situations where the available database is limited to fewer than a few hundred variables.