PaperlayerView exact revision on arXiv

Source-linked AI summary

Coral Grow-out Robotic Assessment System (CGRAS): Scaling Coral Recruit Monitoring Through Robotics and Computer Vision

Dorian Tsai, Scarlett Raine, Emilio Olivastri, Riki Lamont, Andrew Lui, Timothy Morris, Joshua Esplin, Christopher A. Brunner, F. Mikaela Nordborg, Reginald Wardleworth, Garima Samvedi, Karen Jackel, Matthew Dunbabin, Tobias Fischer, Andrea Severati

arXiv:2609.38846v1cs.RO

TL;DR

Frequent monitoring of coral grow-out is labor-intensive, and recruit-scale imaging must work across tank arrays in a humid, saline environment. CGRAS combines robotic image acquisition with computer vision to assess coral recruits; it achieved 96.4% agreement with manual counts for Aken corals and reduced monitoring effort 9.6-fold.

  • Problem

    Grow-out success depends on environmental conditions that are not fully understood, making frequent monitoring necessary, while recruit-scale imaging across tank arrays poses technical challenges in marine conditions.

    Source: §I · p. 2

  • Method

    CGRAS combines robotic high-resolution tile imaging with a processing pipeline that stitches images, detects live corals, and generates health and spatial-distribution outputs.

    Source: §I · p. 2, §III · p. 2

  • Results

    96.4% agreement with manual counts for Aken corals and 9.6-fold reduction in monitoring effort compared with manual operation.

    Source: §VII · p. 8

  • Takeaways & Limitations

    The resulting data enable tank conditions and treatment strategies to be correlated with coral survival dynamics at a resolution manual monitoring cannot match.

    Source: §VII · p. 8

    Counting is limited in densely populated regions where coral colonies are closely spaced or merged.

    Source: §VI · p. 7

Abstract

from arXiv · show

Climate change is the largest threat to coral reefs, with increasing global impacts accelerating the need for scalable reef restoration technologies. Large-scale reef restoration depends on the mass production of corals, such as through coral aquaculture. Coral seeding with recruits grown in aquaculture facilities is a feasible restoration approach, but effective production requires consistent, high-frequency monitoring of tens of thousands of macroscopic (0.5-2mm diameter) recruits, making conventional manual assessment prohibitively labor-intensive. To address this monitoring bottleneck, we introduce the Coral Grow-out Robotic Assessment System (CGRAS) which combines robotic imaging and computer vision to automate data acquisition, perform multi-species detection and counting of corals, and evaluate coral health. CGRAS automatically extracts coral growth, survival and spatial distribution metrics, with the aim of providing timely feedback to operators for optimizing production, grow-out and deployment workflow processes. We demonstrate CGRAS in a large aquaculture facility on standardized coral settlement tiles, reducing the time and labor costs by a factor of 9.6 as compared to manual monitoring, whilst achieving 96.4% agreement for Acropora kenti corals relative to expert counts.

I. INTRODUCTION

Coral restoration increasingly depends on scalable aquaculture, but grow-out is sensitive to environmental conditions and requires frequent monitoring. CGRAS combines robotic imaging with computer vision to automate recruit assessment amid the scale, precision, and marine-environment demands of this work.

I. INTRODUCTION

  • Substantial reef loss is projected by 2050 without effective intervention, and coral restoration is increasingly necessary alongside climate action.Coral bleaching is a stress response involving loss of symbiotic algae, the corals’ primary energy source; it is most commonly associated with heat stress driven by climate change.Source: §I · p. 1
  • Coral seeding uses sexually produced recruits grown in nurseries and transplanted to reefs, while supporting genetic diversity and adaptive potential.It can produce the numbers of corals needed for large-scale restoration interventions.Source: §I · p. 1
  • Grow-out lasts approximately 12 weeks and is environmentally sensitive, while successful cultivation drivers remain incompletely understood and require frequent inspection to tune tank parameters.Tiles are assessed for coral survival and cultivation success under controlled conditions.Source: §I · pp. 1–2
  • Assessing one 280 × 280 mm tile takes 1–3 hours of expert stereoscope work and involves handling that may stress corals.Automation also faces the challenge of resolving millimeter-scale recruits across meter-scale tank arrays in humid, saline conditions without sacrificing image precision or calibration.Source: §I · p. 2
  • CGRAS combines non-invasive, repeatable robotic image acquisition with a processing pipeline that estimates coral counts, health, and spatial distribution over time.The stated aim is to provide timely data for research and real-time decisions.Source: §I · p. 2

II. RELATED WORK

Prior work covers automated coral assessment, benthic imaging, and reef-scale robotics, but does not provide automated image acquisition for tiles in coral production facilities. CGRAS pairs robotic high-resolution imaging with recruit detection for aquaculture grow-out monitoring.

A. Automated Assessment of Early Coral Life Stages

  • Earlier coral-life-stage methods automate assessment under controlled conditions, but rely on imagery already captured individually with a microscope or stereoscope.A hierarchical YOLO-SAM pipeline for recruit segmentation and tracking exceeds 95% mean IoU while reducing analysis time.Source: A. Automated Assessment of Early Coral… · p. 2

B. Imaging Platforms for Benthic Monitoring

C. Robotic and Learning-based Reef Monitoring

  • Reef-scale robotics and learning operate at much coarser spatial resolutions than individual recruits, and larval monitoring systems do not monitor settled recruits.Post-settlement recruits attach to complex surfaces and vary spatially and morphologically during grow-out; CGRAS pairs a robotic arm and high-resolution underwater imaging with recruit detection for facility monitoring.Source: C. Robotic and Learning-based Reef… · p. 2

III. METHOD

CGRAS integrates robotic imaging and image analysis for coral grow-out monitoring. It captures tile imagery, assesses recruit survival over time, and uses heatmaps to inform outplanting choices.

III. METHOD

  • CGRAS combines the Image Acquisition System (IAS), which captures high-resolution tile images, and the Coral Counting & Visualization System (CCVS), which stitches images, detects live corals, and generates spatial heatmaps.The heatmaps show healthy recruit distribution across tile tabs.Source: §III · p. 2
  • During grow-out, CGRAS monitors coral survival for 2 to 12 weeks, with the stated aim of enabling real-time tank-condition optimization.Heatmaps are used to select tiles and tabs for outplanting when recruits reach an appropriate growth stage.Source: Figure 2 · p. 3
  • Calibration tags establish the spatial relationship between the robot arm base and two imaging tanks, allowing tile positions to be localized within the tanks.The system captures both tanks in a single operation.Source: Figure 3 · p. 3

A. Preliminaries: the Coral Grow-out Process

Coral grow-out rears settled larvae on concrete tiles for up to 12 weeks before suitable tabs are prepared for reef deployment. The IAS uses a calibrated robotic platform to image tiles and send the imagery for processing.

A. Preliminaries: the Coral Grow-out Process

  • Larvae settle on preconditioned 280 × 280 mm concrete tiles arranged as a 20 × 20 grid of tabs, then recruits are reared for up to 12 weeks before suitable tabs are selected for deployment.Tiles are assessed for coral size and density; healthy suitable tabs are fitted into seeding devices for reef deployment.Source: A. Preliminaries: the Coral Grow-out… · p. 3

B. Image Acquisition System (IAS)

  • The IAS uses a UR20 robotic arm and an end effector with scientific and depth cameras, lighting, and a Raspberry Pi to capture tile images.The arm is mounted on a mobile platform and protected by a sleeve; the camera is housed in a waterproof flat-port enclosure.Source: B. Image Acquisition System (IAS) · pp. 3–4
  • The imaging setup accommodates up to 25 tiles per tank and can image up to 50 tiles across its two tanks in one session.RFID tags on tile molds associate each imaged tile with its coral-health metrics.Source: B. Image Acquisition System (IAS) · p. 4
  • Operators scan tile RFID tags, calibrate the tank pose using three marked calibration tags, and start an open-loop robotic path that captures overlapping sub-images for stitching.The configurable path captures a 6 × 4 grid of sub-images per tile; the resulting imagery is processed by the CCVS.Source: B. Image Acquisition System (IAS) · p. 4

C. Coral Counting & Visualization System (CCVS)

The CCVS turns acquired imagery into stitched tile representations, coral detections, health trends, and spatial heatmaps. Its patch-based detection and tile-localization steps support tile-level counting and visualization.

C. Coral Counting & Visualization System (CCVS)

  • The CCVS manages data, stitches images, localizes tiles, detects corals, analyzes health trends, and generates spatial heatmaps.Outputs are available through a web interface for visualization and data extraction.Source: C. Coral Counting & Visualization… · p. 4
  • Across the illustrated development sequence, skeletal structures become more pronounced after Week 1 and central polyps darken as symbiotic algae become more abundant.The examples also show mortality after Week 4 for Aken and Week 12 for Amil; Pdae remain smaller and largely transparent until Week 4.Source: Figure 4 · p. 4
  • Black corner spacers provide consistent geometric references for tile localization and template matching after Random Forest-based image binarization.Each batch of 24 sub-images is stitched into an approximately 28,000 × 28,000-pixel tile image; concrete tile corners can be obscured or irregular.Source: C. Coral Counting & Visualization… · p. 4
  • The detector applies sliced inference to overlapping 640 × 640-pixel patches because full-image inference is impractical and down-sampling would lose fine detail.YOLOv8n processes each patch; overlapping detections are merged to avoid duplicate coral counts, with more than 3,000 inference passes required per tile.Source: C. Coral Counting & Visualization… · p. 4

IV. EXPERIMENTAL SETUP

CGRAS was deployed at SeaSim during the 2024 and 2025 spawning events and subsequent grow-out periods. The study collected weekly imagery across selected coral species and used model-assisted annotation to reduce labeling effort.

IV. EXPERIMENTAL SETUP

  • CGRAS datasets came from SeaSim under controlled conditions during the 2024 and 2025 late spring/early summer spawning events and grow-out periods.Source: §IV · p. 5

A. Data Collection

  • The study imaged Aken, Amil, and Pdae weekly, selecting species based on spawning availability, spawn quality, parent-colony numbers, and deployment suitability.Aken and Amil are fast-growing Acropora; Pdae was included as a slower-growing coral with distinct morphology.Source: A. Data Collection · p. 5

B. Data Annotation

  • Corals were labeled as one live-coral class with bounding boxes, and preliminary model predictions from ten manually annotated sub-images were reviewed and corrected in CVAT.After the first 30 manually labeled images averaged 31 minutes per sub-image, model-assisted annotation reduced time for the remaining 110 images per species to 6.5 minutes per sub-image.Source: B. Data Annotation · p. 5

1) 2024 Spawning Dataset:

The study assembled separate spawning-event datasets for 2024 and 2025, then trained species-specific detectors using partitioned data and COCO-pretrained weights. The 2025 Aken model’s mAP50 gains plateaued at about 30 labeled images, although the final model used all 130 labeled images.

1) 2024 Spawning Dataset:

  • The 2024 dataset covered 18 tiles across 12 weeks for Aken, Amil, and Pdae, yielding 3,024 sub-images per species; 140 sub-images per species were annotated for training.Source: 1) 2024 Spawning Dataset · p. 5

2) 2025 Spawning Dataset:

  • The 2025 dataset covered the first four post-settlement weeks for 99 Aken tiles and 83 Amil tiles, with 10 tile images per species annotated for training.The datasets contained 2,376 Aken and 1,992 Amil sub-images.Source: 2) 2025 Spawning Dataset · p. 5
  • For Aken, mAP50 increased rapidly up to approximately 30 labeled training images before plateauing; the final model used all 130 labeled images.The figure reports the learning curve averaged over nine runs and a precision–recall curve for the final model’s alive-coral class.Source: Figure 5 · p. 5

C. Model Training and Inference

  • Each year-and-species dataset was split 0.7/0.15/0.15 for training, validation, and testing, and trained with COCO-pretrained weights until convergence or 1,000 epochs.Training used an NVIDIA H100 GPU with standard data augmentation; inference ran on the CGRAS onboard computer with an Intel 12th Gen i7 CPU and NVIDIA RTX 3060 12GB GPU.Source: C. Model Training and Inference · p. 5

D. Evaluation Metrics

CGRAS evaluation covers imaging and processing operations, detection performance, and count agreement with expert assessments. In the 2025 trial, CGRAS reduced the time for equivalent monitoring by 9.6×, while processing failures were linked mainly to stitching errors.

D. Evaluation Metrics

  • Operational assessment records tiles imaged and processed, non-recoverable errors, and imaging and processing time, with total CGRAS time compared against equivalent manual work.In 2025, CGRAS imaged 183 tiles, with one non-recoverable imaging session interrupted by a power failure.Source: D. Evaluation Metrics · pp. 5–6, A. Tile Image Acquisition · p. 6
  • Detection performance is measured by mAP50 at a 0.3 confidence threshold, while count agreement is summarized using per-tab differences, their mean, and standard deviation.Source: D. Evaluation Metrics · p. 6

A. Tile Image Acquisition

  • The CCVS successfully processed 160 tiles (87.9%); 17 failures (9.3%) were primarily caused by stitching errors, and five (2.7%) by spacer-corner occlusion or cropping.Source: A. Tile Image Acquisition · p. 6
  • 213.5 hours of manual monitoring for 160 tiles compared with 22.2 hours using CGRAS represented a 9.6× reduction in operational time.The production-scale illustration estimates 120,000 manual labor hours for weekly monitoring of 10,000 tiles over 12 weeks, excluding CGRAS capital and maintenance costs.Source: A. Tile Image Acquisition · p. 6

B. Coral Detection

CGRAS detection performance improves with more labeled images, with most gains reached using about 30 images; transfer is stronger between similar Acropora species than to Platygyra daedalea.

1) Minimum Data Required:

  • The training-set experiment used 100 Aken images from the 2024 dataset, added in batches of 10, with fixed 15-image validation and test sets.The images were distributed across the 12-week grow-out period.Source: 1) Minimum Data Required · p. 6
  • Above 80% mAP50 is reached with approximately 30 labeled tile images; using all 100 raises mAP50 to around 85%, only about 5% more for 70 additional images.With all 100 images, precision is 89.1%, recall is 85.9%, and F1-score is 87.5%.Source: 1) Minimum Data Required · p. 6

2) Generalization of Coral Detection Models:

  • The cross-species experiment trained a separate detector for each of three species and tested it on the other species using balanced subsets of 130 images per species.The experiment used the 2024 dataset, which contained all three species; the supplied table caption identifies single-species detectors as columns and test datasets as rows.Source: Table II · p. 6, 2) Generalization of Coral Detection… · p. 7
  • Aken-to-Amil cross-performance exceeds 82.3% mAP50, while transfer to Platygyra daedalea reaches at most 64.1%.The authors suggest a genus-level Acroporidae detector may be practical for CCVS deployment.Source: 2) Generalization of Coral Detection… · p. 7

C. Coral Density Distribution on Tiles

CGRAS visualizes live-coral density across tile tabs and supports selection of tabs for deployment. Repeated imaging also provides survival trends, with manual comparisons showing close average agreement despite substantial absolute count error.

C. Coral Density Distribution on Tiles

  • CGRAS maps live-coral quantities by tab, helping operators identify density hotspots and select tabs for deployment according to recruit density and deployment objectives.The heatmap uses absolute quantities, with brighter colors indicating more live corals.Source: C. Coral Density Distribution on Tiles · p. 7
  • The per-tab count-difference distribution has a mean of 0.15 miscounted corals and a standard deviation of 1.59; most per-tab counts match exactly.The figure presents the distribution of differences between manual and CGRAS counts per tab.Source: C. Coral Density Distribution on Tiles · p. 7, Figure 7 · p. 7

D. Coral Survival Trends

  • Repeated imaging of the same tile four times over 23 days produces temporal survival trends from the live-coral detections.Each plotted point represents the total living corals detected at an assessment.Source: D. Coral Survival Trends · p. 7, Figure 8 · p. 7
  • CGRAS differed from manual counts by 4.4% on average, with a mean absolute error of 193.3 corals per tile.In the example, living-coral estimates fell from 2,944 to 109 over 23 days, highlighting substantial mortality for investigation.Source: D. Coral Survival Trends · p. 7, Figure 8 · p. 7
  • The authors suggest automated trend analysis could underpin future alert systems or predictive survival models.The example's rapid decline could prompt investigation of grow-out or aquaculture-system issues.Source: D. Coral Survival Trends · p. 7

VI. DISCUSSION

CGRAS reduces coral-monitoring effort while providing spatial and temporal survival data at previously impractical scales. Its performance depends on representative training data, and crowded scenes and varied coral appearances remain challenges.

VI. DISCUSSION

  • Detector performance depends on how closely deployment data resemble the training distribution; dissimilar recruit appearances can be confused with background objects.The discussion recommends fine-tuning with representative deployment examples and suggests broader biological states and imaging conditions could improve robustness.Source: §VI · p. 7
  • In densely populated regions, closely spaced or merged colonies can be detected as one instance, contributing to under-counting.The discussion identifies improved separation of individual colonies in crowded scenes as a needed approach.Source: §VI · pp. 7–8
  • The open-loop IAS relies on accurate calibration and repeatable tile positioning; future work includes closed-loop approaches and improved optical configurations.Other proposed directions include alternative robotic architectures, more robust stitching, use of unlabeled data, finer-grained coral classification, and temporal growth analysis.Source: §VI · p. 8

VII. CONCLUSION

  • CGRAS operated reliably through the 2024 and 2025 spawning events, with the IAS enabling large-batch image acquisition with minimal operator intervention.The conclusion describes the platform as integrating robotic imaging and analysis for frequent coral-survival data collection.Source: §VII · p. 8
  • CGRAS achieved 96.4% agreement with manual counts for Aken corals and reduced monitoring effort 9.6-fold compared with manual operation.The system provides visualizations for spatial and temporal survival trends.Source: §VII · p. 8
  • Challenges remain in crowded scenes and robust performance across diverse coral appearances, while extension to additional species and developmental stages is future work.The conclusion says CGRAS is already collecting high-resolution survival data at scales previously impractical.Source: §VII · p. 8
Loading 2609.38846v1…