Assessing index influence in marine mucilage monitoring via interpretable deep learning

Abstract The increasing frequency and severity of environmental disasters, such as the widespread marine mucilage formation observed in the Sea of Marmara in 2021 and 2022, demand advanced monitoring solutions. Traditionally, classification approaches such as random forests and the use of various remote sensing indices have been the methods of choice for monitoring such environmental issues. This study presents a deep learning-based classification approach using Sentinel-2 imagery, combined with various Explainable Artificial Intelligence (XAI) methods, specifically SHapley Additive exPlanations (SHAP), Integrated Gradients, Shapley Additive Global Importance (SAGE), and Local Interpretable Model-Agnostic Explanations (LIME), to evaluate the impact of 22 remote sensing indices on mucilage detection performance. Experimental evaluations on multiple Sentinel-2 acquisitions over the Sea of Marmara demonstrate high classification accuracy and reveal that indices such as NDWI1, AMEI, and AWEI play a dominant role in mucilage-water discrimination. In three-class scenarios, additional indices such as EVI also emerge as influential, particularly for land-related class separability. These insights can guide future index selection and support more informed monitoring and analysis of marine mucilage dynamics, contributing to future environmental management efforts.