Abstract
Coral reefs across the Coral Triangle are experiencing unprecedented thermal stress, and the window for effective local mitigation of bleaching events is often measured in days, not months. Yet the standard reef-monitoring pipeline—collecting benthic imagery via divers or remotely operated vehicles (ROVs) and manually annotating it—routinely takes weeks to complete, a delay we term the manual-annotation bottleneck.
CoralEye is a decentralized, edge-optimized computer vision framework that diagnoses coral health in real time, designed to run entirely offline on consumer-grade hardware. Utilizing an INT8-quantized MobileNetV2 architecture trained on the Benthic Habitat Data (BHD) dataset, CoralEye classifies reef imagery into a three-class triage system: Healthy, Bleached, and Dead.
Three-Class Triage Performance
| Class | Precision | Recall | F1-Score | Support (n) |
|---|---|---|---|---|
| Bleached | 0.95 | 1.00 | 0.97 | 241 |
| Dead | 1.00 | 0.89 | 0.94 | 122 |
| Healthy | 0.99 | 1.00 | 0.99 | 273 |
| Weighted Average | 0.98 | 0.97 | 0.97 | 636 |
Note: The dominant failure mode is Dead-to-Bleached confusion (9.8% of dead images misclassified as bleached due to turf-algae overgrowth resembling pale coral tissue).
Edge-First Design
By moving inference directly to the point of capture, CoralEye eliminates dependency on high-bandwidth connectivity or cloud compute budgets. This bridges the technology gap for under-resourced coastal communities and reef managers across Indonesia and the wider Coral Triangle.
Technical Stack
- Architecture: MobileNetV2 (Inverted residuals & linear bottlenecks)
- Quantization: Post-training INT8 quantization via TensorFlow Lite
- Deployment: Browser-ready TensorFlow.js GraphModel (
model.json+.bin) - Geospatial: Leaflet.js + Turf.js coastline polygon masking