Towards operational benthic habitat monitoring: Integrating Sentinel-2, bathymetry, and environmental drivers in Port Phillip Bay, Australia

Coastal benthic habitats provide vital ecological and economic services but face accelerating degradation. Effective management requires accurate habitat maps and a mechanistic understanding of environmental drivers. However, current machine learning (ML) approaches often produce “black-box” predictions with limited ecological interpretability and fail to integrate the full suite of multi-scale environmental drivers known to structure temperate benthic systems. This study addresses these gaps by developing an integrated remote sensing and explainable artificial intelligence (AI) framework to map temperate benthic habitats and quantify their environmental drivers in Port Phillip Bay, Australia. We combined Sentinel-2 imagery, bathymetry, and oceanographic variables within Google Earth Engine (GEE) to train a Random Forest (RF) classifier, achieving an overall accuracy of 0.78 for eight benthic habitat classes. Crucially, we applied SHapley Additive exPlanations (SHAP) to transform the model into an interpretable ecological tool, revealing non-linear responses and critical thresholds. Elevation (depth proxy) was the primary driver, followed by summer sea surface temperature (SST), wave exposure, and turbidity. We identified distinct thermal niches: macroalgae thrived in warmer waters (>18 °C), while seagrass preferred cooler conditions. Seagrass exhibited a unimodal response to turbidity, peaking at moderate turbidity (∼0.7), and wave exposure defined clear zonation between sheltered algal habitats and exposed reef communities. Notably, water-column correction did not improve accuracy, highlighting the value of integrating bathymetry and terrain derivatives in moderately shallow, well-mixed systems. Our framework delivers not only a contemporary habitat map but also a transparent, reproducible model of ecosystem structure. By bridging remote sensing and mechanistic ecology, this work provides a scalable approach for operational monitoring, vulnerability assessment, and evidence-based conservation in temperate coastal ecosystems. • Sentinel 2 + bathymetry delivered strong benthic classification (OA 0.78; κ 0.75). • Water-column correction reduced accuracy (OA 0.63; κ 0.59) despite higher algal estimates. • Driver models showed high predictive skill (OA 0.885; κ 0.779) from terrain, turbidity, SST, and wave exposure. • SHAP revealed distinct ecological niches: seagrass peaked at moderate turbidity while macroalgae declined with turbidity. • Vegetated habitats aligned with interacting gradients (depth, ∼18 °C SST threshold, wave height, terrain complexity).