Vision-Language Models for Underwater Ecological Monitoring

This work presents SeaQuery, a dual-path visionlanguage system for underwater ecological monitoring that combines segmentation-derived metrics with multimodal reasoning. The system employs two pathways: a segmentation-guided path using a pre-trained SegFormer model to extract quantitative ecological metrics (coral cover, bleaching extent, species counts), and a vision-language path with PaliGemma2-3B for contextual scene understanding. Both pathways feed into a Gemma2 decoder that generates responses by integrating pixel-derived metrics with visual features. Evaluation on CoralScapes and CoralVQA datasets shows the dual-path architecture achieves $\mathbf{9 5. 0 \%}$ accuracy in benthic analysis captioning and $\mathbf{3 9. 8 \%}$ exact match accuracy on visual question answering, outperforming individual pathways ($77.4 \%$ captioning for segmentation alone, $35.0 \% \mathrm{VQA}$ for vision-language alone). The results confirm that the design provides an efficient and reliable method for ecological assessment in marine robotics missions.