Deep learning habitat modeling for moving organisms in rapidly changing estuarine environments: A case of two fishes. (5th June 2020)

• Record Type: Journal Article
• Authors: Guénard, Guillaume; Morin, Jean; Matte, Pascal; Secretan, Yves; Valiquette, Eliane; Mingelbier, Marc
• Is Part Of: Estuarine, coastal and shelf science. Volume 238(2020)
• Publication Date: 2020-06-05
• Subjects: Numerical habitat model -- Artificial neural network -- Tide -- Ecohydraulics -- Two-dimensional hydraulic model -- Numerical terrain model; Estuarine oceanography -- Periodicals; Coasts -- Periodicals; Estuarine biology -- Periodicals; Seashore biology -- Periodicals; Coasts; Estuarine biology; Estuarine oceanography; Seashore biology; Periodicals; 551.461805
• DOI: 10.1016/j.ecss.2020.106713
• Languages: English
• ISSNs: 0272-7714

Abstract

Abstract: Modeling the spatial distribution of mobile organisms under rapidly changing environmental conditions is a challenging endeavor that has to be undertaken whenever the impacts of alterations have to be assessed in dynamic scenarios. We modeled habitat suitability for Lake sturgeon ( Acipenser fulvescens ) and White perch ( Morone americana, both had have been followed by hydro-acoustic telemetry) in an estuarine river section with rapidly changing tidal and hydrodynamic conditions using deep feed-forward Artificial Neural Networks (ANN). Descriptors used were of many types: intrinsic features (species, sexual maturity and gender, and individual character), terrain features, hydraulic and tidal conditions, and time. A set of ANN models with varying degree of complexity, in terms of their number of hidden layers, number of nodes per layers, and regularization parameters, were tried and evaluated using cross-validation. The best model has three layers with 100, 50, and 20 nodes and classified 94.0 % of observations as presence (and 60.6 % of pseudo absences as absences, overall correct classification: 77.3 % ) during the trials. The study highlights that tidal and hydraulic models, coupled with acoustic telemetry and machine learning, can be used to predict the spatial distribution of mobile organisms even in extremely variable ecosystems such as estuaries. Highlights: We modeled the habitat of two fishes: Lake sturgeon and White perch. Habitat is located in the St. Lawrence River estuary, near Île d'Orléans. We used a deep Artificial Neural Network model with many types of descriptors. The model classified 94.0% of observations as presence during cross-validation. Descriptors often displayed non-linearity and varied among the environment.