Stream diatom assemblages as environmental indicators - A cross-regional assessment

Ongoing climatic change and anthropogenic pressure highlight the importance of reliable assessment of the ecological status of freshwaters. Bioindicators are used widely in ecological assessments as biotic assemblages reflect the environmental conditions in current ecosystems. However, the robustness of bioindicators relies on the transferability of indices and models outside the regions they were derived from. We tested the reliability of stream diatom assemblages as indicators of water chemistry and climatic factors in a cross-regional assessment by developing a predictive model with diatom assemblage data from Sweden and using it to model stream conditions in Finland. The inference models and predictions were performed using the Boosted Regression Trees (BRT) method, separately in species and genus levels. The predictive performance of the calibration models in Sweden were good or moderate for both water chemistry and climatic variables, both at species and genus levels. The most important climatic (growing degree days, r(2) = 0.57) and water chemistry variables (pH, r(2) = 0.65; and total phosphorus (TP), r(2) = 0.52) could be inferred from diatom assemblages relatively well. However, predictive performances of the cross-regional models were low (r(2) <= 0.13). Nevertheless, water chemistry variables, conductivity (r(2) = 0.13) and TP (r(2) = 0.12), were predicted the best. The most important diatom indicators for climatic and environmental variables varied between Sweden and Finland. The study showed that diatom assemblages can be robust indicators of water chemistry and climatic variables within the region where the inference models are calibrated. However, their indicator ability may be weak between regions. The reason for the low transferability of the diatom inference models may stem from between-region differences in species realized niches, species pools and/or ecosystems, local adaptation or species identification. Hence, models should only be used with caution in geographical contexts other than the one where they were developed. The calibration data should cover as large geographical area as possible to give reliable predictions when applied in smaller regions.