Scottish Marine and Freshwater Science Vol 6 No 6: Development of a Model for Predicting Large Scale Spatio-Temporal Variability in Juvenile Fish Abundance from Electrofishing Data

Discussion

There are a number of challenges involved in the development of large scale habitat models for predicting juvenile fish densities. These include estimation of densities from electrofishing data and consideration of spatial coherence at regional and network scales. Previous attempts to develop habitat models have generally followed one of three approaches (1) simplified approaches considering only single pass electrofishing data (Godfrey, 2005; Wyatt, 2005) or assuming constant capture probability ( SNIFFER, 2011) (2) approaches involving site specific estimates of capture probability (Lanka et al., 1987; Rosenfeld et al., 2000; Godfrey, 2005) or (3) Hierarchical Bayesian approaches where capture probability and habitat covariates are considered simultaneously (Wyatt 2002, 2003; Rivot et al., 2008). The first approach is clearly misleading since habitat variability in capture probability and abundance could be confounded. The second approach does not make best use of available information, particularly in circumstances where no fish are caught and thus no estimate of capture probability can be obtained (Godfrey, 2005). The final approach often results in models which are time consuming to specify and fit, thus making model comparison and selection difficult.

In this report, a fourth approach was considered where relationships between habitat (characterised at large spatial scales using GIS covariates) and fish abundance were modelled in two stages. Firstly, capture probability was estimated in relation to covariates. Secondly, spatio-temporal variability in density was modelled as a function of fish numbers and covariates, conditioned on estimates of capture probability. This two stage modelling approach and the resulting R package, allowed improved estimates of capture probability, rapid specification, fitting and selection of models. Moreover, the R package allowed fitting of flexible (non-linear) relationships between fish abundance and habitat and for modelling spatially correlated data. Such advanced statistical approaches were not available to investigators for previous analyses of Scottish juvenile salmonid data (Godfrey, 2005).

Many large-scale models of salmon abundance have implicitly ignored the effect of capture probability, focussing on single pass data or data from the first pass of multi-pass fishings (Wyatt, 2005), or assumed a constant capture probability. This includes previous models of Scottish fish densities ( SNIFFER, 2011). These approaches are inconsistent with the findings of a number of studies which have shown that capture probability varies with a range of factors including species and life stage (Borgstrom and Skaala, 1993), sampling protocol, personnel (Niemela et al., 2000) and habitat characteristics (Kennedy and Strange, 1981). This report demonstrated that implicit (use of 1 pass data) or explicit (a single estimated capture probability) assumptions of constant capture probability are inappropriate simplifications. Capture probability varied substantially with organisation (which has a geographic component), with year, with the time of electrofishing and to a lesser degree with other habitat covariates. Failure to correct for differences in capture probability generates spatial and temporal biases and consequently misleading models of fish density.

Models of salmon fry density were successfully fitted to the available electrofishing data. Although some of the covariates ( e.g., Channel Width) could have been affected by data processing errors (particularly snapping to river), there is no reason to assume that these estimates should be biased and as such, these errors are only likely to introduce noise into the observed relationships. Furthermore, many of the important covariates which had a strong effect on abundance are robust to site allocation e.g., Catchment, Altitude, Distance to Sea. A potentially more serious issue involves the introduction of bias associated with partial sampling (single channel threads or edges) of large rivers. A number of these site visits were manually identified in the current project through inspection of individual data records. However, further clarification is required from data providers as to where this has occurred. Fortunately the issue of unrepresentative (potentially biased) sampling appears restricted to large and deep rivers where standard SFCC fishing protocols no longer apply. It is therefore likely that the models presented here provide a reasonable overall description of the spatio-temporal variability in fry abundance and the relative importance of covariates, although these preliminary results should be viewed with caution for sites with a high channel width (Mastermap river widths of ca. 27m).

Wyatt (2005) observed a modal response of 0+ salmon densities to Altitude and UCA about 155m and 123km2 respectively. In this study we observed a non-linear negative relationship with altitude and a non-linear relationship with UCA. Although these results initially appear to contrast, the combined effect of DS (not illustrated in Wyatt, 2005), Altitude and UCA may produce similar modal spatial predictions.

Having accounted for between catchment differences in the distribution of habitat covariates ( e.g., DS, Altitude) the Catchment spatial covariate describes residual variation in the abundance data. This could be considered an average relative catchment performance metric over the monitoring period, so long as observed differences did not reflect uncharacterised natural variability in habitat that affects productivity and sampling was representative of the catchment as a whole. Similarly, the RC covariate indicates within catchment variation in abundance relative to the catchment mean. The combination of Catchment and RC therefore provides the opportunity to examine the performance of areas relative to a mean national expectation.

The concept of reference or optimal condition underpins legislation such as the EU Water Framework Directive, while density expectations are required to interpret electrofishing data and fish population health for local fisheries management. Using the current model a reference condition could be inferred by selecting the best Year, setting Catchment to a median value and setting the effects of pressures (% Urban and potentially % Conifer if this represents commercial forestry) to zero. An alternative approach would involve modelling electrofishing data obtained from strategic stocking of sites over a broad geographic and environmental range or defining reference condition for a time period where there were high adult salmon returns and lower human impact, excluding sites thought to be affected by significant anthropogenic pressure. The latter approach would require historical data from relatively un-impacted (reference) sites. While the consequences of this study are clear in a number of contexts, further work would be required to consider how reference conditions could be used to inform any future development of conservation limits ( e.g., Wyatt and Barnard, 1997).

If the models presented here were to be used to assess the status of juvenile salmonid stocks on an annual basis it would be necessary to establish a strategic national electrofishing program that combined quantitative multi-pass electrofishing with single pass area delimited electrofishing to maximise spatial coverage and minimise uncertainty in the resulting data. The optimal mix of quantitative and single pass fishing would require further consideration, as would the spatial structure of annual census data.