Hybrid machine learning algorithms accurately predict marine ecological communities

Predicting ecological communities is highly challenging but necessary to establish effective conservation and monitoring programs. This study aims to predict the spatial distribution of nematode associations from 25 m to 2500 m water depth over an area of 350,000 km² and understand the major oceanographic processes influencing them. The study considered data from 245 nematode genera and 44 environmental parameters from 100 stations. Data was analyzed by means of a hybrid machine learning (ML) approach, which combines unsupervised and supervised methods. The unsupervised phase detected that the nematodes were geographically structured in six associations, each with representative genera. In the supervised stage, these associations were modeled as a function of the environmental features by five supervised algorithms (Support Vector Machine, Random Forest, k-Nearest Neighbors, Naive Bayes, and Stochastic Gradient Boosting), using 80% of the samples for training, leaving the remaining for testing. Among them, the random forest was the best model with an accuracy of 86.4% in the test portion. The Random Forest (RF) model recognized 8 environmental features as significant in predicting the associations. Depth, the concentration of dissolved oxygen in the water near the bottom, the quality and quantity of phytodetritus, the proportion of coarse sand and carbonate, the sediment skewness, pH, and redox potential were the most important features structuring them. The inference of each association across the whole study area was based on the modeling results of the 8 significant environmental features. This model still correctly classified 90% of test data. Such findings demonstrated that it is possible to infer the spatial distribution of the nematode associations using only a small set of environmental features. The recommendation is thus to permanently monitor these environmental variables and run the ML models. Implementing ML approaches in monitoring programs of benthic systems will increase our prediction capacity, reduce monitoring costs, and, ultimately, support the conservation of marine systems.