Merging computational fluid dynamics and machine learning to reveal animal migration strategies

Abstract

Understanding how migratory animals interact with dynamic physical environments remains a major challenge in migration biology. Interactions between migrants and wind and water currents are often poorly resolved in migration models due to both the lack of high‐resolution environmental data, and a lack of understanding of how migrants respond to fine‐scale structure in the physical environment. Here we develop a generalizable, data‐driven methodology to study the migration of animals through complex physical environments. Our approach combines validated computational fluid dynamic (CFD) modelling with animal tracking data to decompose migratory movements into two components, namely movement caused by physical forcing and movement due to active locomotion. We then use a flexible recurrent neural network model to relate local environmental conditions to locomotion behaviour of the migrating animal, allowing us to predict a migrant's force production, velocity and trajectory over time. We apply this framework to a large dataset containing measured trajectories of migrating Chinook salmon through a section of river in California's Sacramento‐San Joaquin Delta. We show that the model is capable of describing fish migratory movements as a function of local flow variables, and that it is possible to accurately forecast migratory movements on which the model was not trained. After validating our model, we show how our framework can be used to understand how migrants respond to local‐flow conditions, how migratory behaviour changes as overall conditions in the system change and how the energetic cost of migratory movements depends on environmental conditions in space and time. Our framework is flexible and can readily be applied to other species and systems.

Development and Evaluation of a Chinook Salmon Smolt Swimming Behavior Model

“…These results inform understanding of swimming behavior and potential management of juvenile Chinook salmon. For example, the conclusion that smolts were not behaving as passive particles, consistent with previous work [13], is important for managers because it suggests that actions such as non-physical barriers that influence salmon smolt behavior may increase survival by influencing route selection. We did not investigate the drivers of smolt behavior in this paper, however we do suggest that multi-dimensional tracking systems such as that used in this study could be leveraged to disentangle these dynamics.…”

“…Behavioral PTM models and individual-based models can represent fish movement by a wide range of approaches [25]. One approach is to specify instantaneous swimming velocity through time which can vary in response to hydrodynamic or other environmental conditions [13,26]. In some cases, the only data available indicating the distribution of fish through time is trawl data collected at monthly or other coarse time intervals.…”

“…In that case, hypothesized behavior formulations can be evaluated based on the consistency of predicted distribution with catch data from trawls [27]. In contrast, acoustic telemetry data collected at a time interval of several seconds, combined with hydrodynamic modeling, allows estimation of instantaneous swimming behavior of salmon at small spatial scales [13]. The swimming speed can be further analyzed to provide swimming behavior formulations with instantaneous swimming velocities.…”

“…In our analysis, swimming velocities estimated from acoustic telemetry data and three-dimensional hydrodynamic model results were used to infer swimming behavior of Chinook salmon. This study shares aspects of other swimming behavior analyses, such as the use of a hydrodynamic model in swimming speed estimates [13] and route selection studies, including the assumption of surface orientation [14] but we combined more aspects of observed behavior in our behavioral PTM than these previous studies. The most complex behavior formulation was a combination of surface orientation (maintaining a vertical position near the surface), constant rheotaxis and time-varying swimming.…”

Machine learned daily life history classification using low frequency tracking data and automated modelling pipelines: application to North American waterfowl

“…Given the nearly equivalent accuracies produced by the multi-layer perceptron model to the best performing XGBoost model in our empirical example (weighted F1-score = 0.922 vs 0.950, respectively), we expect that more “temporally aware” sequence-dependent prediction frameworks, such as Long Short-Term Memory methods [ 46 ], may improve predictability when class assignments follow in a logical progression [ 50 ]. Similar methods have been used to reveal animal migration strategies [ 47 ] but were not yet implementable within Amazon SageMaker AutoPilot© at the time of our investigation. “Super-learning” or ensemble methods that combine and optimize results from multiple models may likewise improve model performance as they have for accelerometry-based classifications [ 43 ].…”