Marine mucilage mapping with explained deep learning model using water-related spectral indices: a case study of Dardanelles Strait, Turkey

Abstract

Rapid detection and periodic monitoring of mucilage formations are of valuable importance for early warning and mitigation strategies. Recent advances in remote sensing technology offer extraordinary advantages for identifying and monitoring fast-emerging and floating mucilage aggregates. To determine mucilage formations on the water surface, this study utilize a cloud-free Sentinel-2 image acquired on May 21, 2021, when mucilage aggregates were abundantly apparent in the Dardanelles Strait. To investigate the contribution of water-related indices in the delineation of mucilage-covered areas, the normalized difference turbidity index, Normalized difference water inde, and automated mucilage extraction index (AMEI) were employed in a pixel-based convolutional neural network (CNN) model. According to the classification results, the determination of mucilage formations was achieved with about 98% and 95% in terms of overall accuracy, respectively. The results indicated that the AMEI index was the most effective water-related spectral index in distinguishing mucilage formations from clear water, producing the highest classification accuracy. The game theory-based SHapley Additive exPlanations for global explanation and integrated gradients for local explanation methods were also employed to analyze and interpret the intrinsic behavior of the employed CNN models and determine the most effective spectral features in trained models. The results revealed that while the AMEI index was the most influential one, all water indices were relatively more effective compared to the spectral bands. Overall, the findings of this study validate the effectiveness of CNN models for autonomously recognizing and monitoring mucilage formations in the marine environment.