In August 2019, fires swept through the Brazilian Amazon at a rate that overwhelmed on-the-ground monitoring agencies. Brazil's National Institute for Space Research (INPE) reported over 30,000 fire hotspots in a single month. What made the scale visible to the world was not field teams but a constellation of satellites feeding data into AI classification systems β systems that flagged active fire pixels, mapped deforestation polygons, and estimated smoke plumes in near real time.
Remote sensing β gathering information about Earth's surface from aircraft or satellites β predates AI by decades. The U.S. Landsat program began in 1972, producing multispectral imagery that scientists laboriously analyzed by hand. A single Landsat scene covering 185 Γ 185 km could take weeks to classify manually. The European Space Agency's Sentinel constellation, launched from 2014 onward, raised the data volume by orders of magnitude: Sentinel-2 alone generates roughly 1.6 terabytes per day.
Manual analysis at that scale is impossible. AI β particularly convolutional neural networks (CNNs) trained to recognize spectral and spatial patterns β transformed remote sensing from a specialist bottleneck into a continuous, near-global monitoring capability.
Global Forest Watch (GFW), operated by the World Resources Institute, uses Google Earth Engine and ML algorithms to process Landsat and MODIS imagery weekly. Its GLAD (Global Land Analysis and Discovery) alert system flags new deforestation down to 30-metre resolution within days of clearing. By 2023 the system had issued alerts covering more than 1.4 million distinct disturbance events globally.
Satellite sensors capture reflected or emitted radiation across multiple spectral bands β visible light, near-infrared (NIR), shortwave infrared (SWIR), and thermal. Each band encodes different surface properties. Healthy vegetation strongly reflects NIR; burnt scars absorb it. Water absorbs most visible light; urban concrete reflects it broadly.
AI models learn these signatures from labeled training data β human-annotated images where experts have identified forest, cropland, water, urban area, or bare soil. Once trained, the model applies its learned decision boundaries to new imagery at machine speed. A CNN processing a 10,000 Γ 10,000 pixel Sentinel scene produces a land-cover classification map in seconds that would take a skilled analyst days.
The European Union's Copernicus Earth Observation programme provides free, open Sentinel imagery to any researcher or developer. Sentinel-1 (radar, all-weather), Sentinel-2 (multispectral, 10 m resolution), and Sentinel-3 (ocean and land colour) together deliver global coverage every 5β12 days. Platforms like Google Earth Engine, Microsoft Planetary Computer, and ESA's own Sentinel Hub provide cloud APIs that let AI systems query and process petabytes of archived imagery without downloading it.
This combination β dense free imagery plus cloud-native AI processing β has democratized environmental monitoring. Universities in sub-Saharan Africa, NGOs in Southeast Asia, and government agencies in Pacific island nations can now run deforestation or reef-degradation monitoring pipelines that were once the exclusive province of well-funded agencies.
Planet Labs operates a constellation of over 200 small "Dove" satellites that image the entire Earth's landmass daily at 3β5 metre resolution. Norway's International Climate and Forests Initiative (NICFI) struck a deal in 2020 to make Planet's tropical forest imagery free to researchers and NGOs. AI pipelines built on this data can now detect forest clearing events as small as 0.1 hectares within 24 hours β fast enough to support real-time law enforcement responses.
AI remote sensing is not infallible. Cloud cover β persistent in tropical regions β can block optical sensors for weeks. Radar sensors (Sentinel-1) penetrate clouds but require different model architectures. Training data biases mean models calibrated on one ecosystem can fail on another. And satellite revisit times still leave gaps: a 5-day revisit cycle misses disturbances that begin and end between passes.
Active research areas include data fusion (combining optical, radar, and thermal bands to reduce cloud-gap failures), self-supervised learning (training on unlabeled imagery to reduce dependence on expensive annotations), and foundation models for Earth observation β large pre-trained models like IBM and NASA's Prithvi, released in 2023, that can be fine-tuned for flood mapping, crop monitoring, or wildfire risk with minimal task-specific data.
You are advising a small NGO that monitors forest cover in a tropical region. They have access to free Sentinel-2 imagery and want to build an AI-powered early-warning deforestation alert. Use this lab to think through sensor selection, model architecture, training data strategy, and operational limitations.
Somewhere between Tasmania and Antarctica, a cylindrical instrument roughly the size of a bicycle pump drifts at 1,000 metres depth, carried by currents that have never seen a ship. Every ten days it descends to 2,000 metres, then rises slowly to the surface, measuring temperature, salinity, and dissolved oxygen as it climbs. It transmits its data to a satellite, then sinks again. This is an Argo float β one of nearly 4,000 deployed globally since the year 2000. The programme has produced the largest-ever dataset of ocean interior measurements. Machine learning now mines that archive to infer carbon uptake, heat content, and biological productivity at scales and depths no research vessel could sustain.
The ocean is the planet's primary climate buffer. It has absorbed approximately 90% of the excess heat trapped by greenhouse gas emissions since industrialisation, and roughly 25β30% of anthropogenic COβ each year. Yet the mechanisms of this absorption β how carbon sinks vary by season, depth, and biological activity β remain imprecisely known. Uncertainty in ocean carbon uptake is one of the largest sources of uncertainty in climate projections.
Traditional monitoring relied on research vessels: expensive, slow, and unable to maintain continuous coverage. Autonomous platforms β Argo floats, gliders, Wave Gliders, and autonomous underwater vehicles (AUVs) β changed the geometry of data collection. AI is the essential tool for converting the resulting data torrent into scientific and policy-relevant knowledge.
The Monterey Bay Aquarium Research Institute (MBARI) developed biogeochemical Argo (BGC-Argo) floats equipped with oxygen, nitrate, pH, and chlorophyll sensors. Working with the Southern Ocean Carbon and Climate Observations and Modeling (SOCCOM) project, researchers used ML algorithms to reconstruct full-depth carbon chemistry profiles from the floats' sensor readings β filling gaps where direct carbon measurements were impossible. The resulting dataset revealed that the Southern Ocean was a significantly stronger carbon sink in summer than models had predicted, refining global carbon budget estimates.
Sound travels far in water. Passive acoustic monitoring β deploying hydrophones that record ocean sound continuously β captures whale calls, fish choruses, shipping noise, and the crackle of healthy coral reefs. Manually reviewing thousands of hours of audio is impractical; AI audio classifiers make it routine.
The NOAA Passive Acoustic Research Programme has deployed hydrophone arrays across the Pacific, Atlantic, and Arctic. Deep learning models trained on labeled whale call spectrograms can now identify species, individuals (by unique call patterns), and population trends from continuous recordings. The same technology was applied to coral reef health: a healthy reef is acoustically rich; a bleached reef goes acoustually silent. AI-analysed hydrophone recordings now provide reef health indices without any diver entry.
The Allen Coral Atlas, launched in 2020 by Planet Labs, the Vulcan Foundation, and Arizona State University, used AI to produce the first-ever global high-resolution coral reef map from satellite imagery. The atlas classifies benthic habitat (what lives on the sea floor in shallow water) at 5-metre resolution across all tropical reef regions β roughly 348,000 square kilometres of reef area. Monitoring bleaching events β episodes of coral stress caused by elevated sea temperatures β is now conducted by combining NOAA sea surface temperature (SST) alerts with near-real-time Planet imagery fed into classification CNNs.
The 2016 and 2017 bleaching events on the Great Barrier Reef were the first to be comprehensively mapped using AI-assisted analysis. AIMS (Australian Institute of Marine Science) combined drone surveys, manta-tow transect data, and Sentinel-2 imagery with random-forest and CNN classifiers to produce bleaching-severity maps that guided conservation response efforts.
Saildrone autonomous surface vehicles β wind- and solar-powered vessels roughly 7 metres long β carry a suite of meteorological and oceanographic sensors and can remain at sea for months. In 2021 a Saildrone entered the eye of Hurricane Sam in the Atlantic, transmitting live ocean-atmosphere measurements. NOAA is integrating Saildrone data with AI forecast models to improve hurricane intensity prediction, which depends critically on ocean heat available to fuel storm intensification.
A Pacific Island government wants to monitor the health of its exclusive economic zone (EEZ) β coral reefs, fish stocks, and ocean temperature β to support both conservation and fisheries policy. They have limited budget and no research vessel. Use this lab to design a cost-effective AI-powered ocean monitoring system.
In November 2019, Delhi's air quality index exceeded 500 β "hazardous" β on multiple days. The city's official monitoring network comprised fewer than 40 reference-grade stations across 1,500 square kilometres. A resident in Dwarka had no way of knowing whether conditions 2 kilometres to the east β near a major arterial road β were meaningfully worse than at the nearest official monitor. That gap between sparse official data and lived experience is precisely where AI-powered low-cost sensor networks and spatial interpolation models entered the picture.
Reference-grade air quality monitors β instruments that meet EPA or equivalent regulatory standards β cost between $15,000 and $100,000 each and require regular calibration and maintenance. Most cities in the world have fewer than ten; many have none. The WHO estimates that over 90% of the world's population lives in areas where PM2.5 concentrations exceed WHO guidelines, but the majority of those areas have no direct measurement.
Low-cost sensors (LCS) β devices costing $50β$500 β have proliferated since around 2015. They are less accurate individually but can be deployed in dense networks. AI models correct for their known biases (sensitivity to humidity, temperature, and cross-contamination) and fuse their readings with reference station data, satellite aerosol retrievals, and meteorological models to produce spatially continuous pollution maps.
From 2015 to 2019, Google and the Environmental Defense Fund (EDF) mounted air quality sensors on Google Street View cars driving through Oakland, California; London; and other cities. Using GPS-tagged readings from over 50 million measurements and ML spatial models, they produced maps showing that NOβ and black carbon concentrations varied by up to 8Γ within a single city block β a granularity impossible from fixed reference networks. The data directly informed city decisions about truck route restrictions and school siting near highways.
Space-borne instruments now provide global atmospheric composition data. The European Sentinel-5P satellite, carrying the TROPOMI instrument, measures tropospheric concentrations of NOβ, SOβ, CO, methane, ozone, and aerosols at 3.5 Γ 5.5 km resolution with daily global coverage. NASA's TEMPO (Tropospheric Emissions: Monitoring of Pollution) instrument, launched in 2023, provides hourly daytime air quality maps across North America at roughly 2 km resolution β the first satellite designed specifically for pollution monitoring at regulatory-relevant scales.
AI is essential for extracting useful signals from these measurements. Raw satellite radiances must be inverted using radiative transfer models to estimate surface-level pollutant concentrations β a physically complex process. ML models trained on matched satellite-ground-station pairs have improved the accuracy of these inversions substantially, and deep learning models can now disaggregate coarse satellite pixels to street-level estimates using auxiliary data like road networks and building density.
Wildfire smoke has become one of the dominant sources of PM2.5 exposure in western North America and Australia. The 2020 western U.S. wildfire season produced smoke plumes that drove air quality in Portland and Seattle to the worst ever recorded β worse than any industrial city in the world for multiple days.
The National Oceanic and Atmospheric Administration (NOAA) runs the Hazard Mapping System (HMS), which uses AI to detect fire and smoke from GOES-16/17 geostationary satellite imagery updated every 5β30 minutes. ML models then combine HMS fire radiative power estimates with atmospheric dispersion models to forecast smoke PM2.5 concentrations 72 hours ahead at the county level. AirNow, the U.S. public air quality reporting portal, integrated these ML-enhanced forecasts into its public-facing maps starting in 2021.
PurpleAir sells consumer-grade PM2.5 laser particle counters that upload readings to a public map in real time. By 2023 over 30,000 sensors were active globally. The EPA developed an ML correction algorithm β the "US-Wide Correction" β that adjusts raw PurpleAir readings using co-location studies against reference monitors. During the 2020 western wildfire season, this network β processed through EPA's correction model β provided far denser geographic coverage than the federal reference network alone, guiding evacuation and health advisory decisions in small towns far from official monitors.
Methane (CHβ) is 80Γ more potent than COβ over a 20-year timeframe. A major AI monitoring priority is detecting large methane point sources β leaks from oil and gas infrastructure, landfills, and agricultural operations. The Environmental Defense Fund's MethaneSAT satellite, launched in March 2024, uses ML to detect emissions as small as 3 kg/hour from individual facilities. Earlier work using the GHGSat commercial satellite and Sentinel-5P identified "super-emitter" facilities β a small fraction of oil and gas sites responsible for a disproportionate share of total methane emissions β a finding that has since shaped regulatory policy in multiple countries.
A mid-sized city (population 800,000) in Southeast Asia is designing its first comprehensive air quality monitoring network. They want to combine low-cost sensors, satellite data, and a few reference stations to produce hourly neighbourhood-level pollution maps. Budget is limited. Use this lab to work through the design and methodology choices.
In 2013, the University of Minnesota's Snapshot Serengeti project deployed 225 camera traps across 1,125 square kilometres of Tanzanian savanna. Within two years the cameras had captured 1.2 million images β far more than any team of ecologists could classify manually. The project turned to citizen science via Zooniverse to label them. That labelled dataset then became the training corpus for one of the first deep learning wildlife classifiers β a CNN that could identify 48 Serengeti species from camera trap images with accuracy comparable to trained ecologists. That model, and the datasets it spawned, seeded an entire field of AI-powered wildlife monitoring.
Camera traps β motion-triggered cameras placed in the field β are now deployed in the tens of millions globally by conservation organisations, park authorities, and research projects. The data they produce is staggering: the Wildlife Conservation Society estimated in 2022 that camera trap networks globally generate hundreds of millions of images per year. Manual review at that scale is essentially impossible.
MegaDetector, developed by Microsoft AI for Earth and released openly in 2019, is a CNN trained on 3 million labelled camera trap images. It performs a first-pass detection β identifying images containing animals, humans, or vehicles, and filtering blank frames β achieving 97%+ accuracy across diverse ecosystems. This single tool has saved conservation projects an estimated hundreds of thousands of hours of human review time.
Species-level classification requires more specialized models. Wildlife Insights, a Google platform launched in 2019, trains species classifiers on images from partner organisations and provides a cloud API. By 2023 it had processed over 70 million images from 10,000+ camera trap deployments in 50+ countries. The platform uses transfer learning from ImageNet-pretrained CNNs, fine-tuned on region-specific training sets.
The Snow Leopard Trust operates camera trap networks across Central Asia. Snow leopards are notoriously difficult to census by any method. In collaboration with the Snow Leopard Conservancy, researchers trained AI individual-recognition models on the distinctive spot patterns of snow leopard coats β analogous to facial recognition but applied to fur. The system can re-identify individual animals across camera trap stations, enabling population size estimates without requiring physical capture. AI-assisted censuses revised the total estimated population from a rough 4,000β7,500 to a more constrained range supported by systematic survey data.
Passive acoustic monitoring (PAM) extends the principles of NOAA's marine hydrophone work to terrestrial ecosystems. AudioMoth β a Β£50 open-source acoustic recorder developed by the UK's Open Acoustic Devices project β can record continuously for days from a battery. Deployed in arrays, AudioMoth recorders capture bird calls, bat echolocation, insect stridulation, and frog calls β a soundscape that encodes biodiversity information.
BirdNET, developed by the Cornell Lab of Ornithology and Chemnitz University of Technology and released publicly in 2021, is a deep learning model that identifies over 6,000 bird species from audio recordings. It runs on a smartphone or on AudioMoth data. In 2023, a collaboration between BirdNET and the European Union's LIFE programme used it to survey bird diversity across 300 forest sites in Central Europe, producing species richness maps that would have required years of traditional survey work.
iNaturalist, a joint initiative of the California Academy of Sciences and National Geographic Society, has become the world's largest biodiversity observation database. By 2024 it held over 200 million observations of more than 500,000 species, contributed by millions of citizen scientists with smartphone cameras. Its built-in AI identification tool β trained on iNaturalist's own labelled photo dataset β suggests species from photos uploaded by users. The model achieves top-1 accuracy exceeding 90% for the most commonly observed species and over 85% for the top-5 candidates.
The scientific value is substantial. In 2021, analysis of iNaturalist data identified significant range shifts in 30+ invertebrate species in response to climate change β range expansions and contractions that classical survey programmes, with their fixed sites and infrequent visits, had not detected. AI-mediated citizen science is beginning to provide the kind of spatially dense, temporally continuous biodiversity records that systematic ecology needs.
The UK's Centre for Environment, Fisheries and Aquaculture Science (CEFAS) is developing AI pipelines for eDNA metabarcoding of river water samples. A single 500ml water sample, processed through next-generation sequencing and an ML taxonomic classifier, can identify the presence of over 200 freshwater fish, invertebrate, and plant species simultaneously. Deployed across England's river network, the system is projected to replace months of traditional electrofishing surveys and macroinvertebrate kick-sampling with a monitoring approach that is faster, cheaper, and less ecologically disruptive.
A national park authority in a biodiversity hotspot (highland tropical forest, 500 kmΒ²) wants to establish a long-term biodiversity monitoring programme using AI-assisted methods. They want to monitor mammals, birds, amphibians, and freshwater fish. Budget allows for technology investment but not a large permanent field staff. Use this lab to design the monitoring system.