Harmonizing the Deep: A Unified Information Pipeline for Robust Marine Biodiversity Assessment Across Heterogeneous Domains

TL;DR

This work tackles the challenge of robust marine biodiversity assessment across heterogeneous domains by introducing a Unified Information Pipeline that harmonises diverse underwater datasets and evaluates a fixed detector under controlled cross-domain protocols. It distinguishes failure modes driven by scene structure from those due to visual degradation, showing Context Collapse in sparse scenes as a key risk to early-warning detection, and demonstrates that edge-ready deployment on low-cost hardware is feasible with hardware-aware optimisations. The study contributes through a reproducible, instrumented evaluation framework, diagnostic covariates that separate structure from vision, and operational guardrails that promote cross-site comparability and reliable monitoring indicators. The results imply a shift from image-quality improvements toward structure-aware reliability, offering a practical, scalable foundation for long-horizon marine monitoring and future expansion to species-level inference in constrained field settings.

Abstract

Marine biodiversity monitoring requires scalability and reliability across complex underwater environments to support conservation and invasive-species management. Yet existing detection solutions often exhibit a pronounced deployment gap, with performance degrading sharply when transferred to new sites. This work establishes the foundational detection layer for a multi-year invasive species monitoring initiative targeting Arctic and Atlantic marine ecosystems. We address this challenge by developing a Unified Information Pipeline that standardises heterogeneous datasets into a comparable information flow and evaluates a fixed, deployment-relevant detector under controlled cross-domain protocols. Across multiple domains, we find that structural factors, such as scene composition, object density, and contextual redundancy, explain cross-domain performance loss more strongly than visual degradation such as turbidity, with sparse scenes inducing a characteristic "Context Collapse" failure mode. We further validate operational feasibility by benchmarking inference on low-cost edge hardware, showing that runtime optimisation enables practical sampling rates for remote monitoring. The results shift emphasis from image enhancement toward structure-aware reliability, providing a democratised tool for consistent marine ecosystem assessment.

Figures (6)

• Figure 1: Unified annotation distribution. A) Spatial geometry of 300 sampled bounding boxes; B) Centroid density map; C) Width--height distribution illustrating variation in object size across datasets.
• Figure 2: Sample inference image from scenario 1 (DeepFish). Blue bounding boxes indicate predictions, while white boxes denote ground truth.
• Figure 3: Horizontal Correlation Bars for Scenario 1. Performance tracks Structure (FishCount) more than Vision (Turbidity).
• Figure 4: Sample inference image from scenario 2 (Luderick). Blue bounding boxes indicate predictions, while white boxes denote ground truth.
• Figure 5: Horizontal Correlation Bars for Scenario 2.