Evaluate Salmon Redd Habitats Using a Hierarchical Physics-Informed Machine Learning Framework

Abstract:

The upriver bright (URB) fall chinook salmons, the most abundant salmon in the Columbia River Basin, are important to sport, tribal and commercial fisheries. The mainstream Columbia River at 90-km long Hanford Reach provides the only major spawning habitat for URBs. In the past several decades, many field surveys and modeling studies have been conducted at the Hanford Reach to investigate various physical, hydrological, thermal and chemical factors that influence the spawning habitats of the URB. However, results from these conventional logistic regression models tended to over predict habitats, which indicated some influential environmental factors and/or their interactions were overlooked. In this study, we apply Machine Learning (ML) methods to evaluate the contributions of various environmental factors, including both historical surveys and newly acquired datasets, to the redd habitats of URB. A new hierarchical physics-informed ML framework based on Random Forest is developed to reveal the underlying physical mechanisms and interactions among these environmental factors at various levels of complexity. This study will help stakeholders on deciding research and management actions concerning the recovery of fall Chinook salmon in the Columbia River basin and western North America.