Machine learning for modeling North Atlantic right whale presence in the Gulf of Maine

Machine learning for modeling North Atlantic right whale presence in the Gulf of Maine

As a renewable energy specialist writing for the European Future Energy Forum, I’m excited to explore how machine learning techniques are revolutionizing our understanding of the endangered North Atlantic right whale (NARW) and its dynamic habitat in the Gulf of Maine. This region, known for its rapidly warming waters and shifting marine ecosystems, poses critical challenges for the conservation of this iconic species.

Supervised Learning Algorithms

Researchers have leveraged supervised learning algorithms to build predictive models of NARW presence based on a wealth of passive acoustic monitoring (PAM) data. These algorithms, trained on labeled datasets of NARW vocalizations, can automatically detect the characteristic “upcall” sounds made by the whales and classify their presence across vast ocean regions.

One such study, published in the journal Scientific Reports, utilized a generalized baleen whale call detection and classification system to process over 35,600 days of acoustic recordings from the western North Atlantic. By applying a Mahalanobis Distance threshold, the team was able to identify more than 2,500 days with confirmed NARW presence, shedding new light on the species’ year-round distribution and habitat use.

Unsupervised Learning Algorithms

Complementing these supervised approaches, researchers have also employed unsupervised learning algorithms to uncover patterns in NARW movement and behavior that may not be apparent through traditional observation methods. For example, clustering algorithms have been used to identify distinct behavioral modes, such as foraging or migration, based on the whales’ vocalization patterns and dive profiles.

These unsupervised techniques allow scientists to explore the complex, often non-linear relationships between environmental factors, anthropogenic stressors, and NARW habitat selection without relying on pre-defined categories or labels. By revealing previously hidden insights, they can inform more targeted conservation strategies and adaptive management plans.

Deep Learning Models

The latest frontier in NARW presence modeling involves the use of deep learning — a subset of machine learning that leverages artificial neural networks to extract high-level features from raw data. Deep learning models, trained on large datasets of NARW acoustic recordings, visual sightings, and environmental variables, have demonstrated impressive abilities to predict whale presence with a high degree of accuracy.

One such model, developed by researchers at the University of Massachusetts Boston, combines convolutional neural networks to analyze spectrograms of NARW vocalizations with recurrent neural networks to incorporate temporal and spatial context. This powerful hybrid approach has resulted in remarkably precise forecasts of NARW distribution, helping to guide real-time mitigation efforts and proactive management decisions.

North Atlantic Right Whale

The North Atlantic right whale is one of the world’s most endangered large whale species, with a population estimated at just 350 individuals as of 2019. These majestic creatures, known for their distinctive callosities and slow-moving behavior, have faced a long history of threats, from centuries of commercial whaling to modern-day entanglement in fishing gear and vessel strikes.

Habitat and Distribution

Historically, NARWs were found throughout the western North Atlantic, from the southeastern United States to the Canadian Maritimes and even as far north as Greenland and Iceland. However, in recent decades, their distribution has undergone dramatic shifts, with sightings decreasing in some traditional habitats, such as the Bay of Fundy, while increasing in other areas, like the Gulf of St. Lawrence.

The Gulf of Maine, a highly productive ecosystem at the intersection of several major ocean currents, has long been a crucial feeding ground for the NARW. As this region experiences rapid warming and oceanographic changes, understanding the whales’ dynamic presence and habitat use has become increasingly critical for effective conservation efforts.

Conservation Status and Threats

The NARW is currently listed as Endangered under the U.S. Endangered Species Act and Critically Endangered on the IUCN Red List. While the population showed signs of recovery in the 1990s and early 2000s, reaching an estimated 476 individuals in 2010, it has since declined alarmingly, with high mortality rates linked to entanglement in fishing gear and vessel strikes.

Addressing these pressing threats requires accurate, real-time information on NARW movements and habitat use. This is where machine learning-powered presence modeling becomes an invaluable tool, enabling resource managers to proactively implement protective measures, such as dynamic speed limits for ships and temporary fishing gear modifications, in the areas and times when the whales are most likely to be present.

Gulf of Maine Ecosystem

The Gulf of Maine is a dynamic, highly productive marine ecosystem, characterized by a complex interplay of oceanographic currents, nutrient upwelling, and diverse marine life. As one of the fastest-warming bodies of water in the world, this region is experiencing rapid environmental changes that are profoundly impacting its delicate ecological balance.

Oceanographic Conditions

The Gulf of Maine is influenced by the cold, nutrient-rich Labrador Current from the north and the warm, saline Gulf Stream from the south. This unique confluence of water masses, combined with the region’s complex bathymetry and coastal topography, creates a highly variable and dynamic oceanographic environment.

Climate change is now driving significant shifts in these oceanographic patterns, with the Gulf of Maine experiencing a 99% increase in ocean surface temperatures over the past four decades. These dramatic changes are altering the distribution and abundance of key prey species for the NARW, forcing the whales to adapt their foraging strategies and habitat preferences.

Marine Biodiversity and Ecological Dynamics

The Gulf of Maine is home to a rich tapestry of marine life, including iconic species such as the NARW, as well as a diverse array of fish, seabirds, and other cetaceans. This biodiversity is sustained by the region’s highly productive plankton blooms and the intricate food web they support.

As the Gulf of Maine’s ecosystem undergoes rapid transformation, understanding the cascading effects on NARW presence and habitat use is crucial. Machine learning models can help elucidate these complex, nonlinear relationships, identifying critical thresholds and tipping points that may trigger dramatic shifts in NARW distribution and behavior.

Presence Modeling

Data Collection Methods

The foundation of NARW presence modeling lies in the extensive passive acoustic monitoring (PAM) efforts across the western North Atlantic. These long-term data collection campaigns, involving the deployment of hundreds of underwater recorders, have amassed a wealth of acoustic information on the whales’ vocalizations and presence.

In addition to PAM, researchers have integrated a diverse array of other data sources, including visual sightings, satellite telemetry, and environmental variables (e.g., ocean temperature, currents, prey abundance) to build more comprehensive models of NARW habitat use.

Feature Engineering

A key aspect of developing effective machine learning models for NARW presence is the careful selection and engineering of relevant features from these multifaceted datasets. This may involve extracting temporal patterns in vocalization rates, spatial variables describing habitat characteristics, or oceanographic indices that capture the dynamic environmental conditions in the Gulf of Maine.

By leveraging domain expertise and advanced data processing techniques, researchers can transform raw data into insightful features that capture the complex interplay of factors influencing NARW distribution and behavior. This feature engineering process is critical for training robust, generalizable models that can provide reliable predictions.

Model Evaluation Metrics

Evaluating the performance of machine learning models for NARW presence prediction is a nuanced process, as these models must balance accuracy, precision, and sensitivity to the rare, episodic nature of NARW occurrences. Metrics such as area under the receiver operating characteristic (ROC) curve, F1-score, and Matthews correlation coefficient can provide valuable insights into model performance.

Importantly, these evaluation metrics must be considered in the broader context of the conservation and management implications. A model that achieves high overall accuracy but fails to correctly identify critical NARW presence events may be of limited practical use. Ongoing collaboration between researchers and resource managers is essential to ensure that model development and performance assessment are aligned with real-world decision-making needs.

The European Future Energy Forum (​ www.europeanfutureenergyforum.com ​) is at the forefront of showcasing innovative technologies and research that are transforming the renewable energy landscape. By leveraging machine learning to better understand and predict the dynamic presence of the endangered North Atlantic right whale, researchers are contributing to the development of more effective, adaptive conservation strategies – a critical step in protecting this iconic species and the fragile marine ecosystems it inhabits.

Facebook
Twitter
LinkedIn