Deep neutrosophic-embedded fusion-based deep learning for underwater benthic substrate classification

Deep learning based benthic substrate mapping is essential for large scale marine monitoring and automated seafloor analysis. However, model performance degrades in real underwater environments due to noise, mixed substrates, overlapping classes, variations in depth, illumination, and sensing conditions. Conventional deep learning models typically assume clean, well separated classes. When applied to real world classification and mapping tasks, their accuracy decreases, resulting in high misclassifications rates. This highlights their difficulty in effectively representing uncertainty and ambiguity in benthic imagery which leads to increased misclassification. To address this limitation, we propose a Deep Neutrosophic-embedded Fusion framework for uncertainty-aware benthic classification. The proposed approach explicitly models truth, indeterminacy, and falsity within deep feature representations and is integrated convolution-based, state-space-based, and Transformer-based architectures, demonstrating its adaptability across fundamentally different feature learning paradigms. Specifically, we introduce a learnable deep Neutrosophic fusion strategy implemented in intermediate level of the deep learning architectures, that applies uncertainty-aware gating directly at the patch level feature representation stage. Experimental results demonstrate that the proposed Neutrosophic framework consistently improves the performance of convolution-based (ResNet18), state-space-based (Mamba), and Transformer-based (SAEViT) architectures, demonstrating its adaptability across fundamentally different feature learning paradigms. Among the evaluated models, N-SAEViT achieved the largest relative improvement over its corresponding baseline, increasing the test accuracy from 80.72% to 97.60% (20.91%), whereas N-ResNet18 achieved the highest absolute predictive performance, including the highest test accuracy, Macro-F1 score, and Matthews Correlation Coefficient (MCC). N-Mamba also improved its baseline performance, demonstrating the adaptability of the proposed framework across diverse backbone architectures. Correlation analysis further revealed that Hₘₑₐₙ , Edgedensity , and LBPenergy provide complementary color, structural, and texture information, improving class separability under uncertainty and demonstrating the effectiveness of the proposed deep Neutrosophic fusion framework for robust benthic substrate classification.