fruitpunch

Guide · AI & data

AI for Coral Reefs: Monitoring Reefs With Machine Learning

By the FruitPunch research deskUpdated November 15, 20269 min read

In short

Machine learning now does most of the slow work in coral reef monitoring: tools such as CoralNet and ReefCloud label survey photos, with ReefCloud reported at 80 to 90 percent accuracy and 700 times faster than manual analysis. The need is large: NOAA Coral Reef Watch reports that bleaching-level heat stress reached about 84.4% of the world's reef area between January 2023 and September 2025, and the GCRMN found that 14% of the world's coral was lost between 2009 and 2018.

This address once hosted FruitPunch AI, an Eindhoven AI for good community. FruitPunch is a new, independent publication and is not that organization. This guide was written from public sources; the history page explains the change.

AI helps coral reefs mainly by turning the huge volume of reef photos, underwater recordings and satellite images into numbers fast enough for managers to act on. The need is urgent: NOAA's Coral Reef Watch reports that from 1 January 2023 to 30 September 2025, "bleaching-level heat stress has impacted ~84.4% of the world's coral reef area," with mass bleaching documented in at least 83 countries and territories. Machine learning now labels survey photos in platforms such as CoralNet and ReefCloud, flags reef health from sound, maps every shallow reef from orbit and feeds the satellite heat alerts that warn of bleaching. This guide explains each method, the figures behind it and the open tools anyone can use.

Key facts

How fast are coral reefs declining?

The best global answer comes from the Global Coral Reef Monitoring Network (GCRMN), an operational network of the International Coral Reef Initiative. Its 2020 report was built from "almost 2 million observations from more than 12,000 sites in 73 reef-bearing countries" spanning 1978 to 2019. It notes that reefs cover only 0.2% of the seafloor yet support at least 25% of marine species, and that their goods and services are valued at US$2.7 trillion per year.

The executive summary of that report found that "the 1998 event alone killed 8% of the world's coral," that disturbance events between 2009 and 2018 "killed 14% of the world's coral," and that there was "20% more algae on the world's coral reefs in 2019 than in 2010."

The newer GCRMN 2025 report, drawing on contributions from 600 partners in more than 120 countries, territories and economies, extends the record to 2024. Its executive summary reports a 9.5% relative decline in hard coral cover when comparing 2020 to 2024 with the 1980 to 2009 reference period, an 8.9% fall between 2023 and 2024 alone, and global hard coral cover of 25.8% in 2024. Bleaching events have recurred "every four to six years" since 2010, and the report warns that the full impact of the fourth global bleaching event is not yet reflected in its figures.

Measure Figure Source
Reef area hit by bleaching-level heat stress, Jan 2023 to Sep 2025 84.4% NOAA Coral Reef Watch
Same measure, 2014 to 2017 global event 68.2% NOAA Coral Reef Watch
World's coral killed by the 1998 event 8% GCRMN 2020
World's coral lost, 2009 to 2018 14% GCRMN 2020
Relative decline in hard coral cover, 1980 to 2009 vs 2020 to 2024 9.5% GCRMN 2025
Global hard coral cover in 2024 25.8% GCRMN 2025
Relative rise in macroalgal cover since 1980 44.1% (still low at 6.8%) GCRMN 2025

These figures exist because thousands of divers photograph fixed transects year after year. The slow part has always been what happens after the dive, and that is where AI made the biggest difference.

How does AI analyze reef survey photos?

A standard benthic survey produces hundreds of photos of the seafloor. Analysts then classify random points on each image (live coral, algae, sand, rubble and so on) to estimate cover. Machine learning learns those labels from images that experts have already annotated and applies them to the rest.

CoralNet. Started at the University of California San Diego, CoralNet describes itself as "a resource for benthic images analysis" that "deploys deep neural networks which allow fully and semi automated annotation of images." It is free and open source. The site says that "often 50-100% automation can be achieved with minimal reduction in the quality of the final data-product." Its first server came online in fall 2011, and by the November 2016 beta release it held over 300,000 images and a deep learning engine that the team reports "performs on par with human expert annotators."

Independent testing backs this up. A NOAA team (Williams et al., Frontiers in Marine Science, 2019) ran CoralNet on previously analyzed imagery from NOAA's Pacific Reef Assessment and Monitoring Program in the main Hawaiian Islands and American Samoa. Fully automated estimates of site-level coral cover were "highly comparable to those generated by human analysts (Pearson's r > 0.97, and with bias of 1% or less)." Results were weaker for algae, and accuracy improved when labels were simplified from genus to functional group.

ReefCloud. Built by the Australian Institute of Marine Science (AIMS) and co-funded with Australia's Department of Foreign Affairs and Trade, ReefCloud lets teams upload survey photos with location, date, habitat and depth, then uses "machine-learning from trained experts to automate image analysis." When it launched in Palau in April 2022, the International Coral Reef Initiative reported that it analyzes "coral reef composition with 80-90 percent accuracy 700 times faster than traditional manual assessment," and that it partners with the GCRMN.

MERMAID. The Wildlife Conservation Society's MERMAID is a free, open source platform for fish, benthic and bleaching survey data that now includes AI coral classification with CoralNet and can share data with the GCRMN.

The limits are consistent across tools: models do best on common coral groups, need locally annotated images to adapt to a new region, and still need a person to review uncertain labels.

Can AI tell if a reef is healthy by listening to it?

Yes, at least for broad health classes. Healthy reefs crackle with snapping shrimp and fish calls, while degraded reefs are quieter. Passive acoustic recorders can run for weeks, but nobody can listen to all that audio, so machine learning does the sorting.

At one of the world's largest coral restoration programs, a team led by Ben Williams (Ecological Indicators, 2022) compared healthy sites with 90 to 95% live coral cover and degraded sites with 0 to 20%. The best single acoustic index "could only correctly classify 47% of recordings." A machine learning model combining twelve indices reached a 91.7% success rate on individual one-minute recordings. On two restored sites more than two years old, 37 of 38 and 33 of 39 recordings were classified as healthy; on a site restored less than a year earlier, 27 of 33 were classified as degraded.

Sound can also help restoration directly. In a six-week field experiment published in Nature Communications (Gordon et al., 2019), playing healthy reef sound on degraded patch reefs produced "a doubling in overall abundance and 50% greater species richness" of fish compared with silent controls.

The newest step is pretrained audio models. A 2024 study released SurfPerch, trained on ReefSet, an annotated library of reef sounds that is only 2% of the sample count of bird libraries. They found that mixing bird, reef and unrelated audio in pretraining gave the best results for reef sounds. The code sits in Google Research's open Perch repository.

How do satellites and AI map coral reefs?

Most reefs have never been surveyed by divers, so mapping starts from space. The Allen Coral Atlas, run by Arizona State University with partners, publishes global habitat maps that classify reefs into benthic and geomorphic zones, plus a bleaching monitoring system and turbidity detection.

According to the University of Queensland, the Atlas team identified 348,000 square kilometres of shallow coral reefs, down to 20 to 30 metres, and found that about 80,000 square kilometres have a hard bottom where coral tends to grow. The maps used more than 1.5 million samples and 100 trillion pixels from the Sentinel-2 and Planet Dove satellites, with more than 480 contributors, and the data are free through the Atlas and Google Earth Engine. The results were published in Cell Reports Sustainability in 2024.

Satellite maps show where reefs are and what they are made of, but optical imagery cannot see deep or murky water well, and accuracy varies by region.

How do coral bleaching alerts work?

Bleaching alerts are the most widely used reef data product, and they rely on satellite sea surface temperature. NOAA confirmed the fourth global coral bleaching event on 15 April 2024 and noted that Coral Reef Watch heat stress monitoring uses sea surface temperature data "spanning 1985 to the present, from a blend of NOAA and partner satellites."

The core measure is Degree Heating Weeks (DHW). Coral Reef Watch's 5km methodology explains that DHW adds up heat above the bleaching threshold over the prior 12 weeks: "DHWs over 4 °C-weeks have been shown to cause significant coral bleaching; values over 8 °C-weeks have caused severe bleaching and significant mortality." These values drive alert levels from Bleaching Watch up to Bleaching Alert Level 5, which reef managers can receive by email.

This is mostly physics and statistics, not deep learning. Machine learning comes in when field reports and image surveys are used to check how well satellite heat stress predicts real bleaching, and Coral Reef Watch asks reef teams to submit observations, including reports of no bleaching, for that purpose.

What open tools and datasets can people use?

Tool or dataset Run by What you get
CoralNet UC San Diego Free web platform for automated point annotation of reef photos, with an API
ReefCloud AIMS Cloud repository and AI analysis of benthic survey images
MERMAID Wildlife Conservation Society Open source field data collection, CoralNet classification and an R package
Allen Coral Atlas Arizona State University Global reef habitat maps, bleaching monitoring and turbidity layers
Coral Reef Watch 5km products NOAA Daily global sea surface temperature, DHW and bleaching alert data since 1985
gcrmndb_benthos GCRMN Code repository for the GCRMN benthic cover synthetic dataset
SurfPerch and Perch Google Research Pretrained bioacoustic models for reef audio

What did FruitPunch AI's coral reef challenge cover?

This URL once belonged to FruitPunch AI, an Eindhoven based AI for Good community that is not connected to the current FruitPunch site. Its archived AI for Coral Reefs page from December 2021 described a challenge with ReefSupport, an organization building a platform for researchers to analyze ocean surveys automatically. The stated goal was to automate color correction and image segmentation of coral reef photos. This guide is an independent overview of the same topic.

Frequently asked questions

Is AI accurate enough to replace human reef analysts?

For overall coral cover, close to it. A NOAA study found CoralNet's fully automated site-level coral cover matched human analysts with Pearson's r above 0.97 and bias of 1% or less. Accuracy is lower for algae and rare groups, so most programs keep experts reviewing a sample of images.

What is the difference between CoralNet and ReefCloud?

Both are free platforms that use machine learning to classify points on reef survey photos. CoralNet started at UC San Diego in 2011 and is open source. ReefCloud was built by the Australian Institute of Marine Science, launched in 2022 and is linked to the GCRMN's global reporting.

How much of the world's coral has been lost?

The GCRMN found that 14% of the world's coral was lost between 2009 and 2018. Its 2025 report puts global hard coral cover at 25.8% in 2024, after an 8.9% relative fall in a single year, and says the full effect of the fourth global bleaching event is still to come.

How does NOAA predict coral bleaching?

NOAA Coral Reef Watch tracks sea surface temperature from satellites at 5 km resolution and converts heat above each reef's normal maximum into Degree Heating Weeks. Values over 4 °C-weeks are linked to significant bleaching and over 8 °C-weeks to severe bleaching and mortality.

Can sound recordings show whether a reef is recovering?

Yes. In a 2022 study at a large restoration program, a machine learning model classified 91.7% of recordings correctly as healthy or degraded, and it labeled most recordings from restored sites older than two years as healthy. Recorders can run unattended for weeks, so sound can track recovery between diver surveys.

Spotted an error in this guide? Email editorial@fruitpunch.ai

Keep reading