Marine Biology AI: From BRUV Footage to Coral Reef Classification

Why Underwater AI Annotation Is Different

Computer vision annotation for terrestrial imagery has well-established workflows: object detection on clear images, segmentation on high-resolution photographs, action recognition on stable video. Underwater annotation breaks nearly every assumption behind those workflows. Visibility drops within seconds as a fish moves away from the camera. Light attenuation shifts colour balance dramatically — red disappears below 5 metres, changing the visual signature of every object in the frame. Particulate matter in the water column creates motion blur and obscures detail that is essential for species discrimination. And the subject matter — marine fauna — requires species-level classification expertise that almost no general annotator pool possesses.

According to a 2022 benchmarking study of AI-assisted fish identification published in Methods in Ecology and Evolution, deep learning classifiers trained on expert-annotated BRUV footage achieved 87.3% species-level accuracy on Australian reef fish — comparable to trained ichthyologist reviewers on the same footage. Models trained on non-expert-annotated footage from the same surveys achieved 54.1%. The 33-percentage-point gap traces directly to annotation quality, not model architecture or training volume.

The marine AI market in Australia is driven by three overlapping use cases. Conservation monitoring — using AI to process the backlog of survey footage faster than expert ecologists can review manually. Fisheries management — applying species abundance estimates from AI-assisted BRUV analysis to stock assessments and quota decisions. And environmental impact assessment — using AI to monitor reef conditions before and after dredging, marine park rezoning, or climate-driven events such as mass bleaching.

BRUV Footage: What It Is and Why It's Difficult to Annotate

BRUV stands for Baited Remote Underwater Video. A BRUV unit is a steel frame carrying one or two cameras and a bait arm (typically pilchards or chopped oily fish) deployed on the seafloor for 60–90 minutes to attract fish. The resulting footage captures the MaxN metric — the maximum number of individuals of each species simultaneously visible in any single frame — which is used as a non-destructive index of fish abundance.

BRUV annotation is expert-intensive for several compounding reasons. First, the same individual may appear and disappear multiple times during a 90-minute recording; annotators must decide whether a returning fish is the same individual counted previously or a new arrival, which requires visual tracking across frames. Second, similar species in Australian waters require fine-grained visual discrimination: distinguishing sweetlip species requires clear fin and colouration pattern assessment; identifying juvenile versus adult Coral Trout requires size and colouration comparison to reference taxonomic guides; and distinguishing Barramundi from juvenile mulloway requires body depth assessment that varies with camera angle.

Third, water column turbidity varies within a single recording as current changes, fish behaviour stirs the sediment, or bait particles cloud the field of view. Annotators must make species-level identifications from partial or low-visibility frames — a task that requires familiarity with the taxonomic keys, not just visual pattern matching. Our specialist image annotation services apply domain-expert review for exactly these high-stakes identification tasks where generic annotators produce systematic misclassification.

CATAMI Taxonomy: The Australian Marine Classification Standard

CATAMI — the Collaborative and Annotation Tools for Analysis of Marine Imagery standard — is the nationally-consistent classification scheme for Australian marine benthic imagery. Developed through collaboration between the Integrated Marine Observing System (IMOS), the Australian Institute of Marine Science (AIMS), and the Reef Life Survey (RLS) Foundation, CATAMI provides a hierarchical taxonomy covering biota and substrate from major morphological groups (hard coral, soft coral, macroalgae, seagrass, invertebrates, encrusting biota, mobile fauna) down to species or morphological species level for commonly surveyed Australian fauna.

The practical importance of CATAMI for AI training data is national interoperability. Research institutions conducting benthic surveys — from Ningaloo Reef in Western Australia to Lord Howe Island in the Tasman Sea — use CATAMI codes so that their datasets are directly comparable. AI models trained on CATAMI-annotated datasets can therefore be deployed across different survey programmes without retraining. Annotation that does not follow CATAMI creates siloed datasets that cannot be pooled with the broader national monitoring corpus.

For benthic image annotation using CATAMI, annotators place random sample points on each image (typically 25–100 points per image, depending on survey protocol) and assign a CATAMI code to the biota or substrate under each point. This point-count method is used for benthic composition monitoring, coral cover estimation, and macroalgae extent mapping. Polygon segmentation annotation — where each distinct organism or substrate region is outlined as a polygon with a CATAMI label — provides richer spatial data but takes 8–15× longer per image.

Case Study — Great Barrier Reef Coral Bleaching Detection AI

The Great Barrier Reef covers 344,400 km² and is monitored by AIMS's Long-Term Monitoring Programme (LTMP), which has collected annual photographic transect data from 127 reef sites since 1993. The 2022 AIMS annual report documented the fourth mass bleaching event on record, covering more than 91% of surveyed reefs. Processing the resulting imagery backlog — hundreds of thousands of coral colony images requiring expert bleaching severity classification — represented a major constraint on the speed of monitoring reporting.

A university research consortium partnered with AIMS to build a coral bleaching detection AI using the LTMP image archive. The annotation programme annotated 220,000 coral quadrat images with five-class bleaching severity labels — none, pale, partially bleached, mostly bleached, and completely bleached — following the AIMS LTMP bleaching protocol. Annotators were trained coral ecologists with minimum three years of reef survey experience; annotation was conducted using CoralNet with mandatory dual-review for partially-bleached and mostly-bleached categories (the classes with the lowest natural inter-rater agreement).

Annotation programme outcomes: inter-annotator agreement (Cohen's kappa) of κ=0.84 on the final production batch, with κ=0.71 on the pilot batch before guideline iteration addressing the pale/partially-bleached boundary definition. The trained classification model achieved 94.7% accuracy on a held-out test set of 11,000 images reviewed by a senior coral ecologist panel. Model throughput: 14,000 images per hour on standard GPU infrastructure, versus 4–6 images per hour for a single expert ecologist reviewer.

Before/after impact: annual transect data that previously required 18 months of expert ecologist time to classify can now be processed in 3.4 days of AI inference with expert spot-checking on the 8.3% of images the model flags as uncertain. The research team estimates this acceleration enables bleaching severity maps to be available to reef managers and conservation policy teams within weeks of field surveys rather than over a year later — a meaningful improvement for adaptive management response to bleaching events.

Cetacean Survey Annotation and Aerial Marine AI

Australian cetacean monitoring uses aerial line-transect surveys — fixed-wing aircraft or large-format drones flying systematic transects over known cetacean habitat — to estimate population size and distribution. The resulting imagery contains thousands of photographs taken at 150–600 metre altitude, each requiring expert review to detect marine mammals at the surface. AI-assisted detection dramatically increases the throughput of image review, but the training data requirement is substantial.

Annotation for cetacean detection models involves two separate tasks: bounding box annotation of animal locations in each aerial photograph (the detection task), and species classification of each detected animal based on dorsal fin, body size, colouration, and behaviour visible in the image (the identification task). For humpback whales — the most-monitored Australian cetacean species — individual photo-ID annotation (matching fluke patterns across surveys to track individual animals) represents a third annotation task that requires specialist photographic identification training.

The most effective workflow for cetacean aerial survey annotation uses model-assisted annotation: a detection model trained on previously-annotated imagery flags candidate animal locations, and marine biologist annotators confirm, correct, and species-classify each detection. This semi-automated pipeline reduces annotation time by 60–70% versus fully-manual annotation on new survey imagery, while maintaining expert-quality species identification labels. See our overview of video annotation services for the tracking and temporal annotation tasks that apply to drone footage of cetacean groups.

Geospatial and Satellite Marine AI

Beyond in-water and aerial imagery, marine AI is increasingly applied to satellite and airborne remote sensing data. Satellite-based applications include: seagrass extent mapping from multispectral and hyperspectral imagery (Australian seagrass meadows cover approximately 44,000 km² and serve as critical dugong habitat and carbon sinks); kelp forest mapping from Landsat and Sentinel-2 imagery; and marine debris detection in coastal zones. Each satellite application requires annotated training data — polygon labels delineating seagrass, kelp, or debris patches in georeferenced imagery — before AI models can be trained.

The annotation challenge for satellite marine imagery is ground-truth matching: spectral signatures in multispectral imagery do not directly correspond to the visual appearance of organisms in underwater photographs. Annotators labelling satellite imagery as "seagrass" or "coral" must interpret spectral indices — not just image appearance — and validate labels against in-water transect data where available. This requires annotators who understand both remote sensing methodology and marine biology, a combination that is rare in general annotation pools. For broader context on geospatial annotation methodology, our guide on geospatial annotation case studies covers the workflow for remote sensing annotation at scale.

Annotator Requirements and Programme Structure

Minimum annotator qualifications for marine biology AI differ by task. BRUV fish annotation: formal ichthyology or marine biology training (BSc minimum), practical experience identifying Australian marine fauna in the relevant bioregion (tropical Great Barrier Reef species are different from temperate Southern Ocean species), and familiarity with FishBase and the relevant regional fish identification guides. Coral health annotation: reef ecology training with specific experience in coral bleaching survey methodology, and calibration to the specific bleaching scale used (AIMS LTMP, CoralNet BleachWatch, or project-specific). Cetacean aerial annotation: marine mammal biology training and experience in photographic species identification from aerial perspective.

QA structure for marine annotation follows the same dual-review logic as other specialist annotation: a primary annotator completes the label, a senior reviewer confirms or corrects it, and disagreements are escalated to a third expert adjudicator. For novel or ambiguous species — juvenile fish, degraded coral colonies where bleaching state is unclear, partially-visible cetaceans — a three-way consensus requirement is standard. Inter-annotator agreement targets vary by task: coral bleaching κ ≥ 0.75 (reflecting genuine ecological ambiguity in borderline bleaching states); fish species ID accuracy ≥ 85% (reflecting the limits of single-frame species discrimination); cetacean detection recall ≥ 95% (weighted to minimise missed animals rather than minimise false positives).

For research institutions and government conservation agencies sourcing annotation, the parallels with agricultural AI annotation are instructive — both require domain-expert annotators who understand the biology of the subject matter, not just image labelling conventions. Our analysis of agriculture AI annotation covers analogous specialist annotation requirements for crop and livestock monitoring AI.

Related Reading

• What Is Geospatial Annotation and How Is It Used in Mapping and Earth AI?
• What Is Image Annotation and Which Type Does Your Model Need?
• How Is Data Annotation Used in Agriculture AI? Crop, Weed and Yield Case Study

Frequently Asked Questions