Ecological monitoring is critical for conservation efforts, yet these data often feature strong class imbalances which complicate the development of models to predict such events. In their latest research, Michael W. Wade and colleagues propose two modelling frameworks for predicting exceptionally rare aggregatory behaviour of bull and blacktip sharks along the Gulf coast of Texas.

Across many industries, leaders have begun leveraging the immense potential of machine learning to derive powerful insights from large pools of data. Increased processing capabilities and access to statistical programming tools have allowed users to apply these tools to quickly identify complex patterns and make highly accurate predictions. Machine learning is particularly useful in predicting anomalous events, where previous models may have been too conservative to detect these rare cases.

Data where anomalies are present, or zero-inflated data, are common in wildlife conservation and management. This is due to the nature of collecting data via monitoring, which often involves indiscriminate community surveys in multiple study areas. In the Gulf of Mexico, shark populations provide an example through their uncommon – but ecologically important – aggregatory behaviour.

Since the early 1980s, the Texas Parks and Wildlife Department has been monitoring coastal shark populations by conducting routine gillnet surveys along several study sites. Often, these surveys record the presence of just a single shark, if any. In rare cases, several sharks of one or multiple species will be recorded in a single event.

These aggregations, which support feeding and reproduction, serve a significant purpose for not only shark communities, but the broader coastal ecosystem. Therefore, it is critical for wildlife management groups to understand the factors that support and potentially trigger these events in nature.

Attempting to design traditional statistical models to detect patterns between spatial and environmental variables and these aggregation events proves challenging due to the strong class imbalance present between the occurrences of these events and their overwhelmingly more common absence. Imagine this: if an aggregation is observed just 5% of the time, and a model predicts that it never occurs, it will be 95% accurate. While that could be presented as an impressive model, it is effectively useless for those trying to identify future positive cases.

Two machine learning frameworks – gradient boosting and artificial neural networks – provide powerful alternatives to traditional models. Depending on several parameters, both models are adept at extracting deeply complex patterns that contribute to rare events. Interestingly, though, each takes a much different operational approach.

Boosted regression trees, a subclass of gradient boosting, work by first creating a simple branching diagram where the value of a predictor variable determines the most likely outcome. This simple model is scored, and additional branches are iteratively added until a strong, flexible ensemble model exists. These models are not uncommon in ecological research, forming a strong foundation for their development as classifiers of strongly zero-inflated data.

Artificial neural networks exist as a series of strongly interconnected layers of individual nodes. The first layer, called the input layer, contains a single node for each feature, or predictor variable. The values of each input node are then passed through at least one hidden layer, typically containing fewer nodes than the input layer. Along each edge pathway between the layers, an activation function and a bias term are applied to mathematically transform the values.

Finally, those values are passed to the output layer, containing fewer nodes still. A final bias term is applied, backpropagation is optionally applied, and you are left with a predicted outcome. Neural networks, while computationally expensive and difficult to interpret, represent a promising avenue for ecological datasets where many observations are available.

Each of these modelling frameworks, particularly when combined with a random downsampling of negative cases to artificially increase the proportion of positive cases in our training data, demonstrated successful ability to classify cases of aggregatory behaviour. While the neural network performed reasonably well, correctly identifying 82% of positive cases in our validation dataset, the gradient boosting machine demonstrated superior performance and correctly classified more than 87% of aggregation events.

These results demonstrate the effectiveness and flexibility of applying advanced machine learning techniques to analyse ecological monitoring data and suggest that even sparse survey datasets can be used to develop powerful predictive tools. Perhaps more importantly, the extension of these methods could allow managers to better anticipate rare but ecologically important events, allowing monitoring surveys to be performed more efficiently in future cases.

Read the full paper Comparison of two machine learning frameworks for predicting aggregatory behaviour of sharks in Journal of Applied Ecology