Find in Library

There has been increasing use of spatial statistical models to understand and predict river temperature (<i>T</i>_w) from landscape covariates. However, it is not financially or logistically feasible to monitor all rivers and the transferability of such models has not been explored. This paper uses <i>T</i>_w data from four river catchments collected in August 2015 to assess how well spatial regression models predict the maximum 7-day rolling mean of daily maximum <i>T</i>_w (<i>T</i>_{w_max}) within and between catchments. Models were fitted for each catchment separately using (1) landscape covariates only (LS models) and (2) landscape covariates and an air temperature (<i>T</i>_a) metric (LS_<i>T</i>_a models). All the LS models included upstream catchment area and three included a river network smoother (RNS) that accounted for unexplained spatial structure. The LS models transferred reasonably to other catchments, at least when predicting relative levels of <i>T</i>_{w_max}. However, the predictions were biased when mean <i>T</i>_{w_max} differed between catchments. The RNS was needed to characterise and predict finer-scale spatially correlated variation. Because the RNS was unique to each catchment and thus non-transferable, predictions were better within catchments than between catchments. A single model fitted to all catchments found no interactions between the landscape covariates and catchment, suggesting that the landscape relationships were transferable. The LS_<i>T</i>_a models transferred less well, with particularly poor performance when the relationship with the <i>T</i>_a metric was physically implausible or required extrapolation outside the range of the data. A single model fitted to all catchments found catchment-specific relationships between <i>T</i>_{w_max} and the <i>T</i>_a metric, indicating that the <i>T</i>_a metric was not transferable. These findings improve our understanding of the transferability of spatial statistical river temperature models and provide a foundation for developing new approaches for predicting <i>T</i>_w at unmonitored locations across multiple catchments and larger spatial scales.