Dynamic adjustment of training sets ('moving-window' reconstruction) by using transfer functions in paleolimnology - a new approach

摘要

This paper presents a new method (moving-windows) that optimizes diatom-based paleolimnological reconstructions of past environmental conditions from supra-regional training sets. The moving-window method identifies the best number of nearest neighbours (window size) from a merged supra-regional EDDI and local (MV) training set (n = 429) for each fossil diatom assemblage and the best type of transfer function (ML, WA-PLS) based on the error statistic of each transfer function (highest cross-validated R 2, lowest cross-validated average bias, maximum bias and RMSEP). At first we evaluated the moving-window approach by comparing measured TP-values with inferred TP-values using both the moving-window approach and the WA-PLS method. The relative errors of the moving-window approach were not significantly different for 208 samples that had an error <15 mu g/l TP using the WA-PLS method. However, for the remaining 221 samples with errors 15 mu g/l TP using the WA-PLS method, the moving window approach significantly reduced the relative error of the inferred TP levels. Secondly, the moving-window approach was used to reconstruct epilimnetic total phosphorous (TP) for Lake Dudinghausen, Lake Rugensee, Lake Tiefer See and Lake Drewitzer See (Northern Germany) using both the supra-regional EDDI training set and a local training set from Northern Germany (MV training set). The moving-window inferred TP-levels of the four study lakes were compared with published reconstructed TP-values and with inferred TP-values based on the local MV training set. Overall, the moving-window and the published TP-trends agree well with each other. However, the moving-window reconstructions generally indicated lower TP-levels throughout the past similar to 5,000 to 12,000 years, including past maxima. Thus, the moving-window method seems to generate more realistic absolute TP levels due to the optimized window size (highest number of modern analogues, best error statistics). The identification of more realistic absolute historic TP-values is important for the validation of reference conditions for inland waters. This study also demonstrates that a robust local training set may, similar to moving-window training sets, also lead to reliable reconstructions, if the geological settings of the local training set lakes and the study lakes are similar.