Original ArticlesMonitoring riverine thermal regimes on stream networks: Insights into spatial sampling designs from the Snoqualmie River, WA

- Guidance for monitoring riverine thermal regimes beyond summer mean temperature.
- More sites increased predictive precision, but not necessarily predictive accuracy.
- Mean temperatures were easier to model than maximums, minimums, or variability.
- Winter data were less variable and therefore easier to model than summer data.
- Nearby sites with discordant thermal regimes have potential to be highly influential.

Understanding, predicting, and managing the spatiotemporal complexity of stream thermal regimes requires monitoring strategies designed specifically to make inference about spatiotemporal variability on the whole stream network. Moreover, monitoring can be tailored to capture particular facets of this complex thermal landscape that may be important indicators for species and life stages of management concern. We applied spatial stream network models (SSNMs) to an empirical dataset of water temperature from the Snoqualmie River watershed, WA, and use results to provide guidance with respect to necessary sample size, location of new sites, and selection of a modeling approach. As expected, increasing the number of monitoring stations improved both predictive precision and the ability to estimate covariates of stream temperature; however, even relatively small numbers of monitoring stations, n = 20, did an adequate job when well-distributed and when used to build models with only a few covariates. In general, winter data were easier to model and, across seasons, mean temperatures were easier to model than summer maximums, winter minimums, or variance. Adding new sites was advantageous but we did not observe major differences in model performance for particular new site locations. Adding sites from parts of the river network with thermal regimes which differed from the rest of the network, and which were therefore highly influential, improved nearby predictions but reduced model-estimated precision of predictions in the rest of the network. Lastly, using models which accounted for the network-based spatial correlation between observations made it much more likely that estimated prediction confidence intervals covered the true parameter; the exact form of the spatial correlation made little difference. By incorporating spatial structure between observations, SSNMs are particularly valuable for accurate estimation of prediction uncertainty at unmeasured locations. Based on our results, we make the following suggestions for designing water temperature monitoring arrays: (1) make use of pilot data when possible; (2) maintain a distribution of monitors across the stream network (i.e., over space and across the full range of covariates); (3) maintain multiple spatial clusters for more accurately estimating correlation of nearby sites; (4) if sites are to be added, prioritize capturing a range of covariates over adding new tributaries; (5) maintain a sensor array in winter; and (6) expect reduced accuracy and precision when predicting metrics other than means.