2 Making the right choices: understanding the impact of biophysical dispersal modelling
2.5 Influence of modelling decisions on connectivity metrics
2.5.1 Modelling environment
The oceanographic systems were described as either open or closed, based on their geography (see methods). Logically, if the scale of the systems are equal, closed systems are predicted to increase both local retention and settlement success, as closed systems should lead to
decreased rates of larval loss and increased chances of settlement. Self-recruitment is expected to either be lower or equal in a closed system compared to an open system of comparable size, as while more larvae will settle back to their natal patch, there will be more recruits from other patches. Dispersal distance is expected to increase within an open system, as currents can take larvae further off-shore or alongshore. I found self-recruitment and settlement success tended to be higher in a closed system (39% and 12% higher respectively), although the differences were not significant (self-recruitment: t = 1.18, df = 22.7, padjusted = 0.335; settlement success: t = 0.55, df = 126, padjusted = 0.471). The mean local retention in open systems (10.1%) was not significantly higher than in closed system (16.8%; t = -1.60, df = 52, padjusted = 0.229), contrary to the prediction. Dispersal distance was 2.7 times greater in open systems, compared to closed (t = -3.29, df = 65.1, padjusted = 0.006). Only dispersal distance behaved as hypothesised, with the result of local retention being most unexpected. This could be indicative of researchers appropriately choosing the boundaries of the modelled systems based on the biology of the modelled taxa. In this case, even seemingly open systems, may in fact be demographically closed ones. This can be explained by the geographical nature of an open system, potentially allowing more larvae to be dispersed away from natal sites and towards more distant settlement sites or into large oceanic features, there are many more dispersal permutations in open systems.
There is no clear effect of particle-tracking model choice on the resulting connectivity metrics (Appendix A; Figure A1). In general, model choices produced less variability for metrics of local retention and dispersal distance and more variability within measures of settlement success and self-recruitment. Although this trend was not consistent, for example within the choice of particle tracking model, the Connectivity Modelling System (CMS) had relatively consistent predicted values of self-recruitment, but rather variable local retention, settlement success and dispersal distance. Without knowing the specifics of various BDMs, it was expected that there would be inherent variation due to the different implementations. However, any strong trends seen here could be due to the specific code of a model biasing
using more generalised hydrodynamic models (e.g. HYCOM, ROMS, and POM) had the largest ranges amongst the metrics for hydrodynamic models, while more specific models (e.g. MARS-3D and HANSOM) had the least variation (Appendix A; Figure A2). However, for both BDM and hydrodynamic models, the largest variation was seen for the models that were most common amongst studies.
Nested hydrodynamic models (i.e. sub-models with higher resolution to capture coastal dynamics), used in more complex BDMs to capture small scale regional processes, are expected to give more realistic and accurate connectivity patterns. Another less common option is to use unstructured grids, that allow for meshes to be smaller near coastal regions and larger for open water (Sundelöf and Jonsson 2012; Puckett et al. 2014). Following from this, hydrodynamic models with finer resolution should also produce more realistic
connectivity patterns, and model resolution is thought to be a defining characteristic of model variability (Hufnagl et al. 2017). It must be noted that this is not an option for many BDMs, where the choice of hydrodynamic model can be limited for the study area. Finer resolution of the general spatial domain or from using nested models to capture small scale coastal dynamics such as eddies and topographic effects, should in theory influence dispersal (Pineda et al. 2007). If small-scale processes do indeed retain larvae closer to their natal site
(Gawarkiewicz et al. 2007), then it is predicted local retention and settlement success would increase, and dispersal distance would be reduced. Self-recruitment is more complex due to the connectedness of the system — if there are strong transport links with other non-natal source regions, then the effect of the predicted increase in local retention will be potentially counter-balanced by an increase in settlement from non-natal larvae, a process seen in systems where near-shore processes were captured on a finer scale (Teske et al. 2016). Therefore, the simplest prediction being that there will be no change in self-recruitment due to steady increase of local and external settlement. I found only one significant positive predictive relationship between the resolution of the hydrodynamic model and the metrics; local retention (a relationship driven by limited data points, r2 = 0.27, F(1,47) = 17.76, padjusted = 0.000; Figure 2.4 a), the rest were not significant; self-recruitment (r2 = 0.05, F(1,71) = 3.51, padjusted = 0.130; Figure 2.4 b), settlement success (r2 = 0.01, F(1,121) = 0.17, padjusted = 0.680; Figure 2.4 c), or dispersal distance (r2 = 0.03, F(1,62) = 1.83, padjusted = 0.242; Figure 2.4 d). When using nested hydrodynamic models, the self-recruitment significantly reduced (36.5%
this result due to the large unbalanced samples between the groups (only 10 models reporting self-recruitment had nested designs, whereas 65 did not). Similar to model resolution, there was no significant difference in the settlement success or dispersal distance when using nested models, and there were no studies using nested models where I was able to get local retention values. Counter-intuitively to the prediction, local retention was found to significantly
increase as the resolution gets coarser, although it is a relationship driven by two data points at 10 and 15 km (Figure 2.4 a). One possible explanation is that models with coarser
resolutions often are over larger spatial scales and therefore use larger settlement habitats, which could increase local retention. Self-recruitment was found to decrease when nested models were utilised, but did not change with global model resolution, demonstrating the complexity of the metric and its relationship to the connectedness of the region. The results of settlement success and dispersal distance not changing with finer resolution models, also ran counter to predictions. This suggests that coarser models might not change the prediction of dispersal distance, but the settlement metrics might be influenced by the model resolution. The only study within the review to investigate the effect of model resolution found that dispersal distance did not change, but settlement success increased as the model resolution got coarser (Garavelli et al. 2014). Studies outside the scope of the review period have found coarser model resolution decreased the overall settlement success (Huret et al. 2007; Putman and He 2013) or not while not specifically influencing local retention, changes in resolution did vary the dispersal patterns (Kvile et al. 2018). In addition to this increased spatial resolution, the temporal resolution of the models is also important. Finer temporal scales allow for increased accuracy in resolving oceanic circulation. However, finer temporal scales are often compromised for memory requirements in storing large volumes of oceanographic data. Many studies within this review also did not clarify the time scales of both the
Figure 2.4: Comparison of the relationship between three input parameters for biophysical dispersal
models; model resolution (km2; a-d), pelagic larval duration (days; e-h), and settlement competency
window (days; i-l), with four common metrics of connectivity; local retention (%), self-recruitment (%), settlement success (%), and dispersal distance (km). Significant linear regressions are shown with 95% confidence intervals.