A Hybrid Framework for Automated Coral Reef Risk Assessment by Zero-Shot Visual Perception and Fuzzy Cognitive Map Inference

Coral reef ecosystems are among the most biodiverse and ecologically critical environments on Earth, yet they face unprecedented decline due to climate change, ocean warming, and anthropogenic pressure. Effective monitoring at scale remains a major challenge, as conventional survey methods rely heavily on expert labour and in-situ access. In this paper, we propose a hybrid framework for automated coral reef ecological risk assessment that combines zero-shot visual reasoning with causal inference. We employ Gemma 4-E4B, an open-weight multimodal large language model, as a perception layer that extracts twenty-five quantitative ecological features from raw underwater survey images drawn from the Coralscapes dataset, the first general-purpose semantic segmentation benchmark for coral reefs. The extracted feature vector is fed as clamped input activations into a Fuzzy Cognitive Map whose causal weight matrix is initialised from expert biological prior knowledge. Iterative sigmoid dynamics propagate causal influence through thirty-one concept nodes until convergence, yielding a six-dimensional risk profile that quantifies ecosystem health, pollution risk, thermal stress, habitat degradation, recovery potential, and heritage integrity. The entire pipeline operates in a zero-shot regime, requiring no task-specific fine-tuning and no manual annotation, and is deployable on commodity single-GPU hardware. Experiments on 50 test images demonstrate the feasibility of the approach and provide quantitative insight into the ecological risk landscape of reef systems.