The integration of in situ ecological monitoring with high-resolution, multi-source environmental data is increasingly central to spatially explicit conservation planning, particularly for highly mobile marine species whose critical functional areas are difficult to delineate using static tools. Identifying where a species feeds, rather than simply where it occurs, is essential to functionally define critical habitat, since foraging areas include ecological processes very relevant to conservation and the assessment of anthropogenic interactions. A machine learning-based spatio-temporal modeling framework has been developed to predict feeding habitat suitability for the striped dolphin (Stenella coeruleoalba) in the Northern Ionian Sea, Gulf of Taranto (GoT, Central Mediterranean Sea), a biodiversity-rich area subject to multiple human pressures. The model integrates a long-term vessel-based survey dataset (2009–2024; 238 feeding sightings) with daily oceanographic variables derived from Copernicus Marine Service reanalysis products, and geomorphometric descriptors from GEBCO, all interpolated onto a 5×5 km analysis grid. Three machine learning algorithms (Random Forest, XGBoost, and LightGBM) were trained and evaluated under random, spatial, and temporal cross-validation strategies, explicitly testing model robustness to spatial and temporal autocorrelation. Predictions from spatially and temporally validated models were combined into a spatio-temporal ensemble, generating annual feeding suitability maps in 2023 and 2024. Models achieved high and stable discriminative performance (AUC 0.88–0.92) across algorithms and validation schemes, with bathymetric depth, distance from the coastline, and dynamic oceanographic gradients emerging as dominant drivers of feeding suitability. Resulting maps revealed a spatially coherent and temporally persistent feeding hotspot along the westernmost margin of the GoT. This approach opens new possibilities for combining environmental features derived from remote sensing and reanalysis with biological monitoring and machine learning, paving the way for  spatially explicit tool to assess the distribution of critical habitats of species of conservation concerns.