Boosted Regression Trees in Analyzing Altantic Weakfish Seasonal Distribution in the Chesapeake Bay

Atlantic weakfish (Cynoscion regalis) is an important fish species along the east coast of North America ecologically and economically.  It is critical to understand and quantify its temporal and spatial distribution for weakfish stock assessment and management.  We compared the performance of BRTs with commonly used generalized linear models (GLMs) and generalized additive models (GAMs) in analyzing the spatial and temporal distribution of weakfish in the Chesapeake Bay in spring and fall 2002-2013.  The explanatory variables included year, dissolved oxygen concentration, salinity, water temperature, latitude of sampling locations, and water depth in the final models.  The probability distributions of the response variable that we explored for GLMs and GAMs included Poisson, quasi-Possion, negative binomial, and Tweedie. The GAMs and BRTs performed better than the GLMs in terms of model fit when analyzing weakfish distribution on both spring and fall surveys based on the AICs and deviance explained.  The BRTs were superior to the GAMs and GLMs in predicting the weakfish distribution according to the mean squared prediction error from the 10-fold cross-validation. Overall, considering both model fit and prediction accuracy, the BRTs were most suitable for seasonal weakfish distribtution analysis in the Chesapeake Bay.