A hybrid sparse-semantic image classification framework to support marine and coastal monitoring and management

Shi, Ying, Song, Lan, Wu, Luocheng and Li, Huanhuan (2026) A hybrid sparse-semantic image classification framework to support marine and coastal monitoring and management. Ocean & Coastal Management, 277, [108207]. (doi:10.1016/j.ocecoaman.2026.108207).

Abstract

Effective ocean and coastal management increasingly relies on timely, reliable, and interpretable information derived from large volumes of visual data collected by satellites, Unmanned Aerial Vehicles (UAVs), and shipborne sensing systems. Such data underpin a wide range of management and governance functions, including environmental monitoring, maritime surveillance, port operations, and coastal planning. However, the operational use of automated image classification in marine and coastal contexts remains constrained by complex visual conditions and the persistent scarcity of labelled data, limiting its integration into routine decision-support workflows. This study develops a hybrid sparse-semantic image classification framework designed to enhance the robustness and applicability of marine and coastal image understanding in management-oriented settings. The framework integrates structured local representations, which capture geometric and spatial characteristics of marine scenes through a sparsity-driven encoding strategy with locality, non-negativity, and semantic consistency constraints, with global semantic features learned via self-supervised Vision Transformers (ViTs). Specifically, semantic information extracted from the Classification (CLS) token of a pretrained self-Distillation with No Labels (DINO) model is fused with sparse local descriptors to form interpretable and discriminative image representations that remain effective under limited labelled data conditions. Experiments conducted on four public benchmark datasets and three maritime image datasets demonstrate consistent performance improvements over representative state-of-the-art approaches. Beyond classification accuracy, the proposed framework provides a transferable and expandable analytical tool to support marine environmental monitoring, maritime traffic surveillance, and coastal management practices by enhancing upstream image interpretation within monitoring and decision-support workflows in data-constrained marine and coastal contexts.

This record has no associated files available for download.

Contributors

Author: Luocheng Wu

Author: Huanhuan Li

Download statistics