Machine Learning for wildlife and marine environment conservation

The Earth system is facing a systemic ecological crisis where climate change and biodiversity erosion interact with accelerating anthropogenic pressures. Marine ecosystems, strategic for the Blue Economy, have become a crucial testing ground for sustainability policies. Since human well-being and maritime sectors depend on the good environmental status of the sea, systematic observation and rigorous inference from Essential Biodiversity Variables are not optional, but mandatory. Against this background, the rapid expansion of ocean observing systems has outpaced conventional analytical workflows, thus requiring intelligent systems that automate data integration, detect meaningful signals and produce outputs that are both operational and ecologically interpretable. The present PhD dissertation advances that frontier by bringing marine ecology and machine learning into a unified framework designed to acknowledge the inherent complexity of marine systems while yielding management-ready understanding. Specifically, machine learning (Random Forest, LSBoost, RUSBoost, XGBoost, ANN) and deep learning (LSTM) algorithms are used for classification and regression tasks, applied to a set of applications, both for marine wildlife and environmental monitoring and conservation. The designed framework was evaluated through four case studies across the Central – Eastern Mediterranean. The first study addresses odontocete feeding ecology in the Northern Ionian Sea, where existing species distribution models rarely integrate behavior and depth-resolved oceanography in a single predictive workflow. The work filled this gap by developing behavior-aware classifiers (Random Forest and RUSBoost) that directly link oceanographic features to species-specific foraging niches and generate conservation-relevant predictions. The second application concerns loggerhead sea turtle habitat suitability in the Adriatic – Ionian region, where most studies either lack a rigorous design or underused three-dimensional environmental information. The work advances this field by merging satellite tracking data with oceanographic descriptors to resolve coastal–pelagic habitat use under a Random Forest- based pipeline. The third case study focuses on estuarine salinity prediction at the land–sea interface, a domain long dominated by physics-based schemes that can struggle with climate change effects and salinization phenomena. The study addresses literature gap by developing machine learning (Random Forest, LSBoost, ANN) and deep learning models (LSTM networks) that integrate hydrological, tidal, and oceanic forcings while preserving physical coherence and seasonal dynamics for near-term operations. The contribution lies in a comparative learning framework where physics and hybrid models, and data-driven approaches are combined to enhance short-term forecasting capacity. The fourth work highlights how spatial and temporal dependence should guide validation, a factor often neglected in ecological modeling. Using striped dolphin feeding habitat suitability as a case study, models (Random Forest, RUSBoost and XGBoost) were benchmarked under temporal cross-validation versus random partitions, mirroring real deployment. The novelty is the integration of dependence diagnostics into the modeling protocol, turning validation into an ecological design principle. Taken together, this PhD thesis establishes a principled and reusable pathway for predictive ecology that privileges generalization, transparency, and ecological robustness over point performance and potentially equips marine agencies and conservation actors with management-ready outputs under accelerating change.