Statistical species distribution modelling of stream invertebrate and fish communities

Abstract

Freshwater bodies such as lakes and rivers cover only <1% of the Earth’s surface, yet they host approximately 10% of known species. A majority of the human population now lives in close proximity to rivers, and growing human demand for the ecological services that freshwater bodies provide has led to the widespread fragmentation, degradation, and destruction of stream habitats. The natural hydrology and morphology of rivers worldwide has been extensively altered by channelization for flood control and land reclamation, water withdrawals for irrigation, and dams constructed for energy production. Pollution from intensive agriculture and urbanization has degraded water quality, with adverse impacts on freshwater communities such as benthic macroinvertebrates and fish. Benthic macroinvertebrate communities are particularly important to stream ecosystem functioning as they exhibit a relatively high taxonomic and functional diversity (e.g., herbivory, detritivory), are a food source for amphibian, bird, and fish communities, and respond to a wide range of natural and anthropogenic environmental conditions. To better understand how stream communities respond to environmental conditions and inform ecosystem restoration efforts, many industrialized countries have established long-term stream biomonitoring programs that sample stream invertebrate and fish communities at selected geographic sites, in addition to data on environmental conditions. However, these programs have limited human and financial resources that constrain key aspects of biomonitoring design, such as the geographic extent and selection of sites, and the taxonomic level of identification for stream invertebrate taxa. These key aspects of biomonitoring design in turn affect our ability to quantify stream invertebrate community responses to natural and anthropogenic conditions, inform stream management, and improve existing biomonitoring efforts. Multivariate statistical species distribution models (SDMs) have become popular tools to predict species distributions based on environmental conditions. However they have been criticized for weak linkages to ecological theory (e.g., niche theory), limited relevance to management, and for including indirect explanatory variables that have no clear mechanistic effect on species distributions. Recent methodological advances in species distribution modelling have proposed increasingly complex models that reflect ecological hypotheses into the model structure, including hierarchical multi-species and joint SDMs. The key goal of this thesis is to investigate whether the calibration of recently proposed species distribution models to observational data can support inference of how stream invertebrate and fish communities respond to natural and anthropogenic environmental conditions. An additional goal is to explore how key aspects of biomonitoring design (i.e., geographic extent, selection of sites, and taxonomic resolution) affect our ability to learn about potential cause-effect relationships. In pursuing these goals, I began with a comparison of the performance of individual, hierarchical multi-species, and joint species distribution modelling approaches when applied to a dataset of stream invertebrate communities. The results show that, compared to stacked individual models, a relatively simple hierarchical multi-species model improves predictive performance over the whole community while also predicting richness. However, extending the hierarchical multi-species model into a joint model that includes residual correlations between taxa leads to a decline in predictive performance. In addition to improving model performance with minor increases in model complexity, I identify the analytical possibilities of increasingly complex SDMs and conclude that for stream invertebrate communities, a hierarchical multi-species modelling approach improves predictive performance for the community relative to individual models while still offering a parsimonious model. Based on these findings, I selected the hierarchical multi-species distribution model to quantify how stream invertebrate communities respond to natural and anthropogenic conditions. I apply the model to multiple datasets that differ in geographic extent, site selection, and taxonomic resolution to analyze how key aspects of biomonitoring design affect our ability to infer stream invertebrate responses to environmental conditions. The model results show that many stream invertebrate taxa respond to multiple anthropogenic stressors (e.g., indicators of agricultural activities, urban areas), while stream temperature emerged as the most important environmental condition across the datasets. Many genera and species showed stronger and more varied responses to environmental conditions compared to their respective families (particularly among widespread families), although the identifiability of responses among rare taxa remained difficult. Overall, the results indicate that a standardized level of taxonomic identification for taxa among multiple biomonitoring programs would allow future analyses to benefit from an increased sample size, a wider coverage of environmental conditions, and a finer taxonomic resolution. The aim of the third and final study in this thesis is to (i) use conceptual models to identify dominant natural and anthropogenic environmental conditions that have a direct, mechanistic effect on distributions of riverine fish, and (ii) to use the conceptual models to guide the development of statistical SDMs of fish presence-absence and density (i.e., the number of individuals per unit area) observations that incorporate the observation process into the model structures. Using environmental variables for which there is available data, I selected environmental conditions identified in the conceptual model based on the predictive performance of the statistical model. To account for habitat fragmentation by natural and artificial barriers, I developed several spatially-explicit habitat variables that consider the accessible area. The results show that the best predictive presence-absence models for all taxa often include spatially-explicit habitat variables (e.g., the mean stream temperature, morphological state) after accounting for barriers. The predictive performance of the selected presence-absence models for all taxa is reasonably good, except for the widespread Salmo spp. The predictive performance of the proposed fish density models largely exceeded their respective null models, with the model for Salmo spp. achieving the highest predictive performance relative to the models for other taxa and the model for Barbus spp. performing worse than its simpler null model due to overfitting. The differences between the environmental conditions included in the conceptual model and the statistical model can be used to identify potential improvements in the environmental data and to highlight gaps in our prior knowledge about specific taxa, potentially informing future efforts to model fish densities. This thesis has made contributions to the methodology and application of statistical SDMs to freshwater communities by demonstrating how some of the limitations of established species distribution modelling approaches can be overcome to allow additional analytical possibilities of ecological communities, and by applying SDMs to derive practical recommendations to improve stream biomonitoring and management.