Guide · AI & data
AI for Earth: How AI Helps the Climate and the Environment
By the FruitPunch research deskUpdated November 15, 202610 min read
In short
AI helps the environment most where satellites and sensors produce more data than people can inspect. ECMWF put its AI forecasting system AIFS into operations on 25 February 2025 with roughly 1,000 times less energy per forecast, Google's AI flood model covers 100 countries and 700 million people, and a machine learning pipeline found 2,974 methane plumes in one year of TROPOMI satellite data. The cost side is real too: the IEA puts data centres at about 1.5% of world electricity in 2024, or 415 TWh, rising to about 945 TWh by 2030.
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 the climate and the environment mainly by reading huge streams of satellite, weather and sensor data faster and more cheaply than people or older software can. The clearest case is weather forecasting: on 25 February 2025 the European Centre for Medium-Range Weather Forecasts took its Artificial Intelligence Forecasting System (AIFS) into operations, reporting gains of up to 20% on some measures such as tropical cyclone tracks and "a reduction of approximately 1,000 times in energy use for making a forecast." Similar methods now map forest loss, flag methane leaks, detect oil slicks at sea, forecast river floods in places without gauges and identify animals in camera trap photos. AI also has its own footprint, mostly through the electricity used by data centres. This guide covers both sides with named projects and open datasets you can use.
Key facts
- 25 February 2025: ECMWF's AIFS became operational, running side by side with its physics-based IFS (ECMWF).
- 90% of 1,380 verification targets: GraphCast beat the most accurate operational deterministic system on that share (Lam et al., arXiv 2212.12794).
- 97.2% of 1,320 targets: GenCast showed greater skill than ECMWF's ensemble forecast ENS (Price et al., Nature, 2024).
- 100 countries, 700 million people: coverage of Google's AI riverine flood forecasts after the November 2024 expansion (Google).
- 4.3 million hectares of tropical primary forest lost in 2025, down 36% from 2024, according to satellite data from the University of Maryland GLAD lab (WRI Global Forest Review).
- 415 TWh, about 1.5% of world electricity: data centre use in 2024, projected to reach about 945 TWh by 2030 (IEA).
How does AI improve weather forecasts and climate models?
Traditional numerical weather prediction solves the equations of the atmosphere on supercomputers. Machine learning models instead learn how the weather evolves from decades of past data, mostly the ERA5 reanalysis, which ECMWF describes as covering "the past 8 decades" with data "available from 1940 onwards."
Google DeepMind's GraphCast was the first widely cited result. Its paper says it "predicts hundreds of weather variables, over 10 days at 0.25 degree resolution globally, in under one minute" and "significantly outperforms the most accurate operational deterministic systems on 90% of 1380 verification targets" (Lam et al.). The follow-up, GenCast, produces probabilistic ensemble forecasts. In Nature in December 2024, its authors reported "greater skill than ENS on 97.2% of 1,320 targets," with better predictions of extreme weather, tropical cyclone tracks and wind power production.
ECMWF then built its own model and made it operational. According to ECMWF, the AIFS uses the same starting conditions as the physics model, built from "around 60 million quality-controlled observations," and runs every six hours. It has a grid spacing of 28 km, compared with 9 km for the physics-based IFS. So AI models are fast and cheap to run but still depend on physics-based starting conditions and are coarser than the best physics model.
Benchmarks such as WeatherBench 2, a benchmark for "global, medium-range (1-14 day) weather forecasting," let researchers compare models on the same terms.
| Model | Built by | What it does | Published result |
|---|---|---|---|
| GraphCast | Google DeepMind | 10-day deterministic forecast at 0.25 degrees | Beat the top operational system on 90% of 1,380 targets (arXiv) |
| GenCast | Google DeepMind | Probabilistic ensemble forecast | Greater skill than ENS on 97.2% of 1,320 targets (Nature) |
| AIFS | ECMWF | Operational open model, 28 km grid | Gains up to 20% on some measures, about 1,000 times less energy per forecast (ECMWF) |
| Flood forecasting model | Google Research | Riverine flood forecasts in ungauged basins | Up to five-day lead time as reliable as current global nowcasts (Nature) |
How is AI used to track deforestation from space?
Forest monitoring is the oldest large-scale use of machine learning on satellite images. The University of Maryland's GLAD lab produces global tree cover maps at 30 m resolution from Landsat composites, and these feed Global Forest Watch, the free platform run by the World Resources Institute.
The latest annual numbers show both progress and scale. The WRI Global Forest Review reports that "the world still lost 4.3 million hectares of tropical primary forest in 2025, equivalent to more than 11 football (soccer) fields of forest per minute." Tropical primary forest loss fell 36% from 2024, a year of record fires, and Brazil saw a 42% reduction. Fires caused 42% of the 25.5 million hectares of global tree cover loss in 2025.
The main limits are tropical cloud cover, which optical satellites cannot see through, and the gap between tree cover and natural forest, since plantations can look like forest from above.
Can satellites and AI find methane leaks?
Yes, and this is one of the fastest moving areas. Methane plumes are invisible to the eye, but instruments such as TROPOMI on the Copernicus Sentinel-5P satellite measure methane concentrations across the whole planet every day. The problem is volume. A study in Atmospheric Chemistry and Physics explains that TROPOMI "produces millions of observations each day," which "makes manual inspection infeasible." The authors used a convolutional neural network to find plume-like shapes and a second classifier to reject artifacts. Applied to 2021 data, it detected 2,974 plumes with a mean estimated source rate of 44 tonnes per hour, coming from 94 persistent clusters and hundreds of transient sources.
Newer missions target smaller leaks. MethaneSAT, developed by the Environmental Defense Fund, notes that "small, widespread sources account for at least 70% of methane emissions from oil and gas production" and publishes its data for free, including on Google Earth Engine. Carbon Mapper runs a public data portal with "observations of thousands of methane and CO2 plumes" from aircraft and satellites. At the national level, Climate TRACE uses AI and machine learning to analyze data from "more than 300 satellites, more than 11,000 sensors" and other sources to build an open inventory of emissions by source; its site lists data from more than 2.7 million emissions sources.
How does AI detect oil spills and illegal fishing at sea?
Cerulean, run by the nonprofit SkyTruth, is "a free, open-source platform that automatically detects oil slicks from vessels and offshore oil platforms across the ocean using satellite imagery and AI." A SkyTruth global analysis published on 23 September 2026 estimates that oil slicks cover 1.58 million km² of ocean surface each year. Greenpeace Romania used its detections to map 226 verified slicks in Romanian Black Sea waters. Cerulean can also display Sentinel-1 radar satellite coverage.
The same approach exposes ships that do not broadcast their position. A 2024 Nature study by Global Fishing Watch and partners combined satellite radar, optical imagery and deep learning to map vessel activity in coastal waters from 2017 to 2021. It found that 72 to 76% of the world's industrial fishing vessels are not publicly tracked, and that 21 to 30% of transport and energy vessel activity is missing from public tracking systems.
Can AI forecast floods where there are no river gauges?
Most rivers in poorer countries have no gauges, which is why flood warnings have been weakest where floods hit hardest. Google Research trained a model across many basins so it can forecast rivers it has never seen measured. Its Nature paper (Nearing et al., 2024) shows that AI forecasting "achieves reliability in predicting extreme riverine events in ungauged watersheds at up to a five-day lead time that is similar to or better than the reliability of nowcasts" from the Copernicus Emergency Management Service's global system.
The forecasts are public through Flood Hub. In November 2024 Google expanded coverage to 100 countries (up from 80) in areas where 700 million people live (up from 460 million), and announced an improved model with the same accuracy at a seven-day lead time, an upcoming API and an expert layer on Flood Hub with close to 250,000 forecast points in 150 countries.
How does AI help protect wildlife and biodiversity?
Biodiversity monitoring produces more photos and audio than any team can review by hand, so AI does the first pass.
- Camera traps. MegaDetector, an open model under the MIT license, finds animals, people and vehicles in camera trap images so researchers can discard empty frames. Wildlife Insights, a shared platform where anyone can upload camera trap images for automatic identification, says its model "is trained to recognize 1,295 species and 237 classes higher in the taxonomic tree" and only marks images blank when it is confident.
- Sound. BirdNET "uses deep learning to recognize over 11,000 species globally" and offers open source tools, including small models that run on low-power field devices.
- Habitat. Satellite foundation models such as NASA and IBM's Prithvi, released openly on Hugging Face, are aimed at uses NASA lists as "tracking changes in land use, monitoring natural disasters, and predicting crop yields."
Most projects still need a person to check uncertain predictions, especially for rare species.
How much energy and carbon does AI itself use?
AI's environmental cost comes mainly from electricity for training and running models in data centres, plus the hardware and the water used for cooling. The IEA's Energy and AI report estimates that "data centres accounted for around 1.5% of the world's electricity consumption in 2024, or 415 terawatt-hours (TWh)," with the United States at 45%, China at 25% and Europe at 15%. It projects consumption will "more than double to around 945 TWh by 2030," which is "slightly more than Japan's total electricity consumption today," and names AI as the most important driver.
Per-model estimates are rarer. A study of the 176 billion parameter BLOOM language model (Luccioni et al.) estimated its final training emitted about 24.7 tonnes of CO2 equivalent counting only the power used during training, and 50.5 tonnes when equipment manufacturing and other processes are included. Against that, ECMWF's AI model cut the energy per forecast by about 1,000 times, so the net effect depends on the use and on how clean the electricity is.
What open datasets and tools can people use?
| Resource | Type | What you get |
|---|---|---|
| ERA5 | Reanalysis dataset | Hourly global weather and climate data from 1940, free via the Copernicus Climate Data Store |
| WeatherBench 2 | Benchmark | Common scores and data for comparing 1 to 14 day forecast models |
| GraphCast code and weights | Model | Released alongside the paper |
| ECMWF AIFS | Operational open model | Medium-range forecasts from an ML model |
| Global Forest Watch tree cover loss | Satellite dataset | Annual 30 m tree cover loss from the UMD GLAD lab |
| Cerulean | Platform and API | Automated oil slick detections from radar imagery |
| Carbon Mapper data portal | Dataset | Methane and CO2 plume observations |
| Climate TRACE | Emissions inventory | Source-level greenhouse gas estimates |
| MegaDetector | Model | Animal, person and vehicle detection in camera trap images |
| BirdNET | Model and app | Bird sound identification for over 11,000 species |
| Prithvi | Foundation model | Pretrained on NASA Harmonized Landsat and Sentinel-2 data |
What was FruitPunch AI's "AI for Earth" challenge?
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 Earth page from December 2021 described a challenge run with the European Space Agency using multispectral data from the Sentinel-2 satellite pair to detect human impact such as wildfires, deforestation, marine litter and melting glaciers. The page said teams of over 30 data specialists would work through September and October. This guide is an independent overview of the same topic.
Frequently asked questions
Is AI better than traditional weather forecasting?
On many standard scores, yes for medium-range forecasts. GraphCast beat the leading deterministic system on 90% of 1,380 targets and GenCast beat ECMWF's ensemble on 97.2% of 1,320. ECMWF still runs its AIFS alongside its physics model, which supplies the starting conditions.
Which satellites are used for AI environmental monitoring?
Projects in this guide rely heavily on free missions: Copernicus Sentinel-1 radar (oil slicks), Sentinel-2 optical imagery, Sentinel-5P TROPOMI (methane) and Landsat (tree cover). Specialized missions such as MethaneSAT and the satellites used by Carbon Mapper add higher resolution greenhouse gas data.
Does AI cause more emissions than it saves?
There is no single answer. The IEA expects data centre electricity use to grow from 415 TWh in 2024 to about 945 TWh by 2030, while specific applications such as ECMWF's AIFS use about 1,000 times less energy per forecast than before. The net effect depends on the use and the electricity mix.
How accurate is AI at detecting deforestation?
Annual maps from the UMD GLAD lab are the standard global reference, and they report 4.3 million hectares of tropical primary forest loss in 2025. Accuracy is lower under persistent cloud and where plantations or regrowth look like natural forest, so field checks remain part of the process.
Can anyone use these AI tools for free?
Many of them, yes. ERA5, Global Forest Watch, Cerulean, Climate TRACE, MegaDetector, BirdNET and Prithvi are all free to access, and Google's flood forecasts are public through Flood Hub, with an API announced for partners and researchers.
Data behind this guide
- Generative AI Statistics53 sourced figures
Spotted an error in this guide? Email editorial@fruitpunch.ai