Standing watch over Hong Kong’s coral reefs
ReefBeacon fuses NOAA satellite heat-stress data with 20+ years of AFCD coral-survey baselines to alert conservation teams before a marine heatwave bleaches the reef.
- 9
- marine parks & reserves monitored
- 84
- reef-building coral species (more than the Caribbean)
- 80+
- AFCD survey sites (since 1997)
- Daily
- NOAA 5km satellite SST updates
Why it matters
Day-scale threat, month-scale data
Hong Kong has 84 species of reef-building coral — more biodiversity than the Caribbean. But as the ocean warms, bleaching is shifting from occasional to a recurring threat, and existing monitoring cadence can’t keep up.
Invisible heatwaves
The Hong Kong Observatory publishes sea-surface temperature (SST) monthly, but marine heatwaves are day-scale sudden events. A monthly reading literally cannot see a multi-day killer heatwave.
Missing acute events
AFCD has run a coral survey since 1997 across 80+ sites — but only once a year, in summer, by volunteer divers. A priceless baseline that nevertheless misses acute bleaching between surveys.
Bleaching is here
A notable bleaching event hit Hong Kong in 2017; the 2024 global mass bleaching (the 4th on record) struck local corals again. Under climate change, the need for early warning has never been more urgent.
The solution
Deterministic-first, AI-augmented
A daily, traceable, human-free alerting loop. Every threshold decision is made by deterministic code — the LLM never does the math — while AI handles orchestration and narrative.
Two-level risk grading
Based on the internationally-agreed Degree Heating Week (DHW) thresholds: DHW > 4 is Watch, DHW > 8 is Bleaching Risk. Grading is fixed by code and not hand-editable.
Delivered in 30 minutes
Once a threshold is crossed, a structured alert reaches the coordinator within 30 minutes via email + in-app notification, with delivery receipts and an audit trail.
Every alert is traceable
Each alert is bound to a raw-data snapshot and a model/threshold version. Any conclusion can be traced back to the NOAA data and the rule that produced it.
How it works
Two loops, year-round
- 01
Daily ingest
Reads NOAA CRW satellite SST / HotSpot / Degree Heating Week (DHW) products, reconciled to zero error against the NOAA portal.
- 02
Deterministic grading
TypeScript applies the DHW > 4 / > 8 thresholds and grades each park’s heat stress.
- 03
AI orchestration
Gemini Flash orchestrates the pipeline and writes the natural-language assessment — it never performs arithmetic.
- 04
Multi-channel push
A threshold breach generates an alert delivered via email and in-app to the coordinator, with a delivery receipt.
Annual offline loop
- 01
Survey photos
Underwater photos from the AFCD annual survey are ingested.
- 02
Multimodal grading
Gemini Pro grades bleaching severity (mild / moderate / severe) with a confidence score; low-confidence results are auto-flagged for human review.
- 03
Event library + RAG
Grades feed the bleaching-event library; Gemini Embedding RAG answers historical questions with citations (2017 / 2024 events).
- 04
Quarterly report
Gemini Pro long-context auto-generates the previous quarter’s bleaching trend report, with data window and source attribution.
Core features
From daily alerts to annual insight
Daily heatwave detection
Degree Heating Week (DHW) daily grading driven by NOAA CRW satellite products, producing each park’s heat-stress tier within 24 hours.
Automatic two-tier alerts
A threshold breach auto-generates a structured alert delivered via email + in-app within 30 minutes, with delivery receipts and audit logs.
Multimodal photo grading
Gemini Pro grades survey underwater photos for bleaching (mild / moderate / severe) with a confidence score; low-confidence auto-routed for human review.
Cited historical Q&A
RAG over the 2017 / 2024 bleaching events — answers must carry traceable citations, and honestly say “insufficient cases” rather than fabricate.
Quarterly trend reports
Auto-generates the previous quarter’s bleaching trend report: heat-stress curve, grade distribution, per-park breakdown, with source attribution.
Consistent colour-graded map
Alert tiers (Watch / High-risk / Normal) shown consistently in three places: email subject, system label, and map layer colour and text.
Coverage
9 marine protected areas, on constant watch
ReefBeacon monitors all 9 of Hong Kong’s marine parks and reserves, spanning the major coral waters from east to west.
- Hoi Ha Wan Marine Park海下灣海岸公園2.56 km²
- Yan Chau Tong Marine Park印洲塘海岸公園6.41 km²
- Tung Ping Chau Marine Park東平洲海岸公園3.79 km²
- Sha Chau & Lung Kwu Chau Marine Park沙洲及龍鼓洲海岸公園13.45 km²
- The Brothers Marine Park大小磨刀海岸公園10.01 km²
- Southwest Lantau Marine Park大嶼山西南海岸公園6.57 km²
- South Lantau Marine Park南大嶼海岸公園23.7 km²
- North Lantau Marine Park北大嶼海岸公園24.75 km²
- Cape D'Aguilar Marine Reserve鶴咀海岸保護區0.2 km² †
† Cape D’Aguilar is a Marine Reserve whose boundary is an approximate placeholder (honestly labelled); the other 8 use official polygon boundaries.
Built for
Four kinds of users
Conservation coordinators
AFCD Marine Conservation: receive heatwave alerts, decide whether to trigger conservation plans, plan the annual survey.
Marine researchers
HKU Institute of Marine Ecology / CUHK: access long-term ecological data for bleaching attribution research.
Conservation NGOs
WWF-Hong Kong / Ocean Park Conservation Foundation HK: data for advocacy and public education.
Park managers
AFCD Marine Parks authority: decide whether to issue temporary conservation notices and adjust park management.
Why it’s trustworthy
Explainable, traceable, degradable
- Deterministic-first: every number and threshold is decided by code; the LLM never touches arithmetic.
- Every alert is bound to a raw-data snapshot and a model/threshold version — conclusions are traceable.
- When AI is unavailable it degrades gracefully to deterministic/template paths; the system never stops.
- Credentials are scrubbed before any event reaches Sentry; sensitive data never leaves the system.
Built on