Diatom-based models for inferring past water chemistry in western Ugandan crater lakes

Diatom surface sediment samples and corresponding water chemistry were collected from 56 lakes across a natural conductivity gradient in western Uganda (reflecting a regional climatic gradient of effective moisture) to explore factors controlling diatom distribution. Here we develop a regional training set from these crater lakes to test the hypothesis that this approach, by providing more appropriate and closer analogues, can improve the accuracy of palaeo-conductivity reconstructions, and so environmental inferences in these lake systems compared to larger training sets. We compare this output to models based on larger, but geographically and limnologically diverse training sets, using the European Diatom Database Initiative (EDDI) database. The relationships between water chemistry and diatom distributions were explored using canonical correspondence analysis (CCA) and partial CCA. Variance partitioning indicated that conductivity accounted for a significant and independent portion of this variation. A transfer function was developed for conductivity (rjack2 = 0.74). Prediction errors, estimated using jack-knifing, are low for the conductivity model (0.256 log10 units). The resulting model was applied to a sedimentary sequence from Lake Kasenda, western Uganda. Comparison of conductivity reconstructions using the Ugandan crater lake training set and the East Africa training set (EDDI) highlighted a number of differences in the optima of key diatom taxa, which lead to differences in reconstructed values and could lead to misinterpretation of the fossil record. This study highlights issues of how far transfer functions based on continental-scale lake datasets such as the EDDI pan-African models should be used and the benefits that may be obtained from regional training sets.