Abstract
California’s native salmonid populations are declining, as evident by the 2008 fishing closures on one historically abundant species, Chinook Salmon (Oncorhynchus tshawytscha). One major impact on the spring-run of Chinook Salmon within the Central Valley has been the damming of natal rivers, severely limiting available spawning habitat. Additionally, many of the streams used by spring-run Chinook Salmon lack extensive habitat data, such as substrate composition, velocity, depth, and woody debris availability, and specific factors limiting spawning habitat suitability are poorly understood. Bayesian Networks are one modeling method that could help to understand these systems and direct restoration efforts toward the most limiting factors within a watershed. These networks are capable of incorporating quantitative data (e.g., derived from empirical studies, literature review, and publicly available spatial data) and qualitative data (e.g., expert elicitation), making them a powerful tool for decision making in data-poor environments. Bayesian Networks are also easily updatable as new empirical data become available. I constructed a Bayesian Network for a Northern California stream, Deer Creek in Tehama County, to provide a useful tool for guiding restoration of spring-run Chinook Salmon spawning habitat. I developed the network using habitat variables thought to be indicators of habitat quality, including stream slope, average width, mean minimum coniferous cover from above, soil type, water year type, and potential existence of a partial barrier downstream. I used the Norsys Netica software to establish the Bayesian Network, and applied this network to each subreach (defined in this context as a riffle-pool stream segment) to determine the suitability of each subreach for Chinook Salmon spawning. Probability of redd (spawning nest site) presence over 50% was used to indicate good habitat suitability for spawning. Redd data was split into two independent sets. I used redd data from one 6 km reach to fit the model (i.e., develop conditional probabilities by back calculating from known outcomes), and used redd data from a second 6 km reach for prediction and comparison with the empirical data for purposes of model validation. I used two types of model validation. I conducted a sensitivity analysis on the network, to determine the influence of each independent variable and determine whether it had an unexpected or disproportionate effect on the outcome. I also conducted an ANOVA comparing redd densities from subreaches predicted to be good spawning habitat against those predicted to be poor spawning habitat by the network, to assess if there was a statistically significant difference between the two. Of the four scenarios I modeled with the network, three exhibited significantly higher redd densities in subreaches designated as good spawning habitat according to probability of redd occurrence (National Hydrography Dataset streamline under dry conditions, traced streamline under dry conditions, and traced streamline under non-dry conditions). The National Hydrography Dataset (NHD) streamline under non-dry conditions overestimated likelihood of redd presence. This was likely due to an exaggerated effect of mean minimum coniferous cover from above within the NHD model. My results, particularly using the traced streamline network, indicate that Bayesian Networks can be used to predict habitat use and prioritize spawning habitat restoration for Chinook Salmon in a data-poor northern California watershed.