Explaining Changes in Fish Community Biomass Using Pressure Indicators: Comparison of Data Analysis Methods and Regional Results

Abstract

This thesis focuses on assessing relationships among ecological indicators, including identifying pressures that best explain changes in the fish community of two Northwest Atlantic ecosystems. The Grand Bank experienced complex ecological changes over three decades, including a rapid collapse and partial recovery of fish biomass, and I synthesized fish community, environmental, and human indicators that reflect these changes. I first used this suite to demonstrate that relationships among fish functional groups changed after the collapse, identify a representative subset of pressure indicators, show the response to pressures varies over different time scales, and illustrate that a common conceptual framework can be misleading. Next, I compared multivariate linear regression (MLR) and non-linear neural networks (NN) for modelling the biomasses of six fish functional groups using fishing and environmental pressures, identified the most influential pressures, and assessed the effect of different delay types and lengths. In contrast to MLR, the delays had negligible impact on NN fit, which illustrates the powerful ability of NN to extract patterns from data. However, MLR generally had better fit than simple 1-hidden node NN ensembles. Both approaches showed that top-down and bottom-up pressures are influential, and that the most influential pressures changed after the collapse. A preliminary assessment of NN predictive power showed that future efforts should continue investigating NN forecast ability. Another case study applying these approaches to the Georges Bank fish community supported these main conclusions. Different pressures were influential for each region, highlighting the need for ecosystem-specific indicator sets. My thesis contributes to knowledge of past and present dynamics of these ecosystems and can potentially inform ecosystem based fisheries management approaches. I recommend MLR models over NN for this application because they are easier to construct and interpret, although NN may be able to provide complementary information through forecasts. Finally, I discuss implications of my findings and suggest future work to build on this research.