Causal learning-driven semantic segmentation for robust coral health status identification

Global warming is accelerating the degradation of coral reef ecosystems, making accurate monitoring of coral reef health status crucial for their protection and restoration. Traditional coral reef remote sensing monitoring primarily relies on satellite or aerial observations, which provide broad spatial coverage but lack the fine-grained capability needed to capture the detailed structure and health status of individual coral colonies. In contrast, underwater photography utilizes close-range, high-resolution image-based observation, which can be considered a non-traditional form of remote sensing, to enable fine-grained assessment of corals with varying health status at pixel level. In this context, underwater image semantic segmentation plays a vital role by extracting discriminative visual features from complex underwater imaging scenes and enabling the automated classification and identification of different coral health status, based on expert-annotated labels. This semantic information can then be used to derive corresponding ecological indicators. While deep learning-based coral image segmentation methods have been proven effective for underwater coral remote sensing monitoring tasks, challenges remain regarding their generalization ability across diverse monitoring scenarios. These challenges stem from shifts in coral image data distributions and the inherent data-driven nature of deep learning models. In this study, we introduce causal learning into coral image segmentation for the first time and propose CDNet, a novel causal-driven semantic segmentation framework designed to robustly identify multiple coral health states — live, dead, and bleached — from imagery in complex and dynamic underwater environments. Specifically, we introduce a Causal Decorrelation Module to reduce spurious correlations within irrelevant features, ensuring that the network can focus on the intrinsic causal features of different coral health status. Additionally, an Enhanced Feature Aggregation Module is proposed to improve the model’s ability to capture multi-scale details and complex boundaries. Extensive experiments demonstrate that CDNet achieves consistently high segmentation performance, with an average mF1 score exceeding 60% across datasets from diverse temporal and spatial domains. Compare to state-of-the-art methods, its mIoU improves by 4.3% to 40%. Moreover, CDNet maintains accurate and consistent segmentation performance under simulated scenarios reflecting practical underwater coral remote sensing monitoring challenges (including internal geometric transformations, variations in external environments, and different contextual dependencies), as well as on diverse real-world underwater coral datasets. Our proposed method provides a reliable and scalable solution for accurate and rapid spatiotemporal monitoring of coral reefs, offering practical value for long-term conservation and climate resilience of coral reefs.