CSLICS: Coral Spawn & Larvae Imaging System

• CSLICS is an automated, submersible imaging platform that employs modular camera modules and deep learning to continuously monitor coral spawn and larvae.
• The system integrates surface and sub-surface imaging modes with YOLOv8 detectors to accurately quantify developmental stages and dramatically reduce manual labor.
• It incorporates advanced digital image processing techniques, such as digital image correlation and motion magnification, to capture fine-scale larval dynamics.

The Coral Spawn and Larvae Imaging Camera System (CSLICS) is an automated, submersible imaging platform developed to address the bottleneck of labor-intensive coral spawn monitoring in aquaculture for reef restoration. CSLICS provides continuous, non-invasive quantification of coral spawn and larvae developmental stages using low-cost camera hardware and deep learning object detectors trained via human-in-the-loop annotation. The system enables rapid scaling of coral aquaculture pipelines, accurate spawn counting, and fine-grained monitoring of fertilization success, thereby contributing to the expansion of restoration efforts to mitigate climate change threats to coral reef ecosystems (Tsai et al., 22 Sep 2025).

1. System Engineering and Imaging Hardware

CSLICS integrates modular camera modules mounted on larval rearing tanks to enable continuous image acquisition across embryogenesis stages. The platform employs a Raspberry Pi High Quality Camera with a Sony IMX477 sensor coupled to a Pimoroni microscope lens, achieving microscopic imaging of coral spawn ranging 150–500 μm with the sensor positioned approximately 5 cm above the water surface. To ensure operational durability in marine environments, the system is housed in enclosures rated to at least IP68, safeguarding against saltwater ingress. Power and data connectivity utilize a Power-over-Ethernet (PoE) board (15 W at 48 V) connected to an 8-port PoE switch, enabling remote power cycling and simplifying installation logistics.

CSLICS is designed to operate in two modes, facilitating adaptive imaging across spawn development:

• Surface Mode ($t_0$–$t_1$): Deployed during the initial 12–24 hours after tank stocking when spawn are buoyant and fragile. Continuous imaging at the water surface enables real-time estimation of fertilization success.
• Sub-Surface Mode ($t_1$–$t_2$): Initiated with the onset of embryogenesis and increased aeration (12–24 hours post-fertilization), tracking developing larvae within the water column away from surface perturbation.

This mode architecture enables comprehensive lifecycle monitoring and avoids mechanical disturbance of the spawn or larvae.

2. Dataset Acquisition and Annotation Strategies

Data acquisition leverages continuous imaging (one frame every 10 seconds, approximately 360 images per hour), with deployment of three CSLICS units per tank to ensure spatial and temporal coverage across developmental stages. The annotated datasets encompass:

• Surface Data: Approximately 2,000 images labeled via a human-in-the-loop, semi-supervised pipeline. Manual annotation of an initial set of 100 images trains a preliminary object detector that produces pseudo-labels for subsequent images, which are then reviewed and corrected. Classes include egg (unfertilized), first cleavage (indicating fertilization), two-cell, four-to-eight-cell, advanced embryogenesis, and "damaged" (excluded from fertilization success metrics).
• Sub-Surface Data: Later-stage larvae images are annotated manually (using CVAT) for single-class detection, focusing exclusively on in-focus larvae for reliable counting.

This annotation architecture is optimized for dataset expansion with limited expert intervention, facilitating robust detector training in diverse image conditions.

3. Deep Learning-Based Detection and Quantification

CSLICS leverages two YOLOv8 object detectors, selected for optimal inference speed and accuracy on embedded hardware. The YOLOv8x 640p variant achieves image inference in approximately 2 seconds per frame on the Raspberry Pi platform.

• Surface Detector: Multi-class detector distinguishing embryogenesis stages supports real-time computation of fertilization success ratios.
• Sub-Surface Detector: Single-class detector enumerates in-focus larvae, maximizing counting accuracy in subsurface imagery.

The computation of fertilization success utilizes class counts in the formula:

$$f = \frac{n_c + n_2 + n_4 + n_a}{n_e + n_c + n_2 + n_4 + n_a}$$

where $n_e$ = eggs, $n_c$ = first cleavage, $n_2$ = two-cell, $n_4$ = four-to-eight-cell, and $n_a$ = advanced developmental stages. This metric provides bounded ($0\leq f\leq1$) real-time indicators of spawn viability over rolling temporal windows.

Detector	Classes	F1 Score	Imaging Mode
Surface YOLOv8	5 (developmental)	82.4%	Surface
Sub-Surface YOLOv8	1 (larva)	65.3%	Sub-surface

4. Experimental Validation and Operational Impact

CSLICS validation occurred during mass spawning events at laboratory aquaculture facilities, directly comparing automated counts with manual sampling.

• Surface Operation: The multi-class detector achieved F1 scores exceeding 72.8% for all developmental classes (82.4% overall for viable developmental stages). Continuous monitoring delivered high-resolution rolling mean fertilization success over three-minute windows, clearly differentiating successful and unsuccessful events.
• Sub-Surface Operation: The single-class detector reached an F1 score of 65.3% ([email protected] = 64.0%). CSLICS sub-surface imaging counts demonstrated close alignment with manual sampling (e.g., RMSE of 45,670 corals for a culture), sufficient for rearing tank management and restoration planning.

Labor reduction is substantial: manual counting of 60 tanks (288 per tank per event) amounts to 5,760 labor hours per spawning event. CSLICS reduced this requirement by approximately 5,720 hours per event, providing significant operational efficiency and scalability.

5. Integration of Subtle Motion Quantification Techniques

Beyond static counting, CSLICS can incorporate digital image processing methods from recent research on coral motion (Li et al., 2020) for enhanced behavioral monitoring.

• Digital Image Correlation (DIC): Provides pixel-wise displacement (horizontal $u_x$, vertical $u_y$) and strain maps ($E_{xx}$, $E_{yy}$, $E_{xy}$), enabling sub-micron motion detection.
• Optical Flow (Lucas–Kanade): Yields pixel-scale velocity field (magnitude and direction), revealing dynamic movement of larvae and tissue margins.
• Phase-Based Motion Magnification: Decomposes frames using a complex steerable pyramid, amplifies phase variations in chosen frequency bands by a factor (up to 75$\times$), making imperceptible movements visually explicit:

$$I_m(x, y) = A_{\omega,0}(x, y) \cdot \exp\left\{i\left[\phi_{\omega,e}(x, y) + \alpha \cdot \Delta\phi_{\omega,e}(x, y)\right]\right\}$$

This approach enhances analysis of behavioral rhythms and spatial mode shapes of spawn and larvae, facilitating refined studies of attachment, dispersion, and responses to environmental stimuli.

A plausible implication is that motion magnification and DIC could be tailored for CSLICS to resolve larvae micro-movements and tissue dynamics not detectable through object counts alone.

6. Influence of Illumination and Behavioral Monitoring

Experimental evidence demonstrates significant behavioral variation of coral tissue and larvae under varying illumination regimes (Li et al., 2020). Upward tissue motion dominates under white/daylight, while downward movement intensifies under blue/night conditions, with more pronounced movement at night.

This suggests CSLICS deployments should employ adaptive illumination protocols to capture authentic behavioral rhythms and spawning events. Incorporating programmable lighting conditions during monitoring enhances detection of fertilization and larval dispersal behaviors, supporting ecological research that seeks to link spawn movement patterns to environmental factors.

7. Significance for Reef Restoration and Future Directions

CSLICS delivers accurate, real-time spawn and larvae quantification with substantial labor savings and minimal biological disturbance. Its integration of object detection, semi-supervised annotation strategies, and advanced image processing addresses longstanding challenges in coral aquaculture monitoring and supports the upscaling of restoration pipelines (Tsai et al., 22 Sep 2025).

Potential future enhancements include:

• Adoption of higher frame-rate imaging for improved temporal resolution.
• Real-time deployment of motion magnification and dynamic mode shape analysis from digital image processing literature to track behavioral events.
• Advanced noise reduction algorithms for subsurface imaging and adaptive parameterization of motion quantification for specific larval movement scales.

The combination of automated, non-invasive monitoring with robust computational vision establishes CSLICS as a blueprint for empirical coral spawn and larvae research, providing scalable solutions as restoration initiatives intensify to counteract climate change impacts on marine ecosystems.