Microsoft has begun bankrolling an ambitious conservation experiment: funding TealWaters’ Wetland Intrinsic Potential tool through its AI for Good Lab to reveal wetlands that have vanished from conventional maps for centuries. The project, centered on Washington state, blends lidar data, multispectral imagery, and machine learning to produce maps at 1–5 meter resolution that expose small, forested, and ephemeral wetlands long invisible to planners and regulators.
A 2023 reconstruction published in Nature estimated that 21% of the world’s inland wetlands—roughly 3.4 million square kilometers—have been lost since 1700, with the steepest declines concentrated in Europe, the United States, and China. These ecosystems sequester vast amounts of carbon, filter water, buffer floods, and sustain biodiversity, but they continue to disappear under pressure from agriculture, urbanization, and climate-driven sea-level rise. A 2025 global assessment warned that if current rates persist, the economic toll from lost wetland services could reach trillions of dollars by mid-century.
TealWaters’ core innovation is the Wetland Intrinsic Potential layer, a spatial probability surface that fuses high-resolution aerial photographs, digital elevation models, lidar-derived microtopography, hydrological indicators, historical drainage records, and machine learning models trained on field-verified observations. The result is a map that not only detects obvious marshes but also picks up subtle terrain signatures hidden under forest canopy—places that legacy remote sensing or coarse inventory methods routinely overlook.
Microsoft’s role is foundational. The AI for Good Lab is providing grant funding, Azure cloud credits, and development support to scale the tool across Washington’s diverse biomes, from temperate rainforests to coastal marshes and inland peatlands. The award explicitly backs a high-resolution statewide wetland inventory, a map of carbon-dense peatlands, and cloud-deployed models that estimate carbon storage and hydrologic drivers. For Microsoft, the partnership fits a broader pattern of deploying AI for environmental monitoring—biodiversity trackers, disaster assessment models, and now wetland intelligence—while also showcasing Azure’s muscle in processing petabytes of geospatial data.
Washington was chosen as a pilot because its varied landscapes stress-test the algorithms’ ability to generalize. In Snohomish County, the Tulalip Tribes are already using TealWaters’ maps to prioritize restoration of huckleberry patches, salmon-rearing habitats, and floodplain connections. Tribal agency staff overlay the WIP surface with cultural-use layers, rapidly identifying hotspots where a replanted wetland serves both ecological and ceremonial purposes. The Washington Department of Ecology and county planners are trialing the maps in environmental review and climate adaptation planning, embedding them into permitting workflows that have historically relied on decades-old, coarse-resolution federal inventories.
TealWaters frames its technology as an amplifier of local expertise, not a substitute. The approach lets communities that lack expensive wetland consultants make science-informed decisions. Still, any tool that claims to “find hidden wetlands” must clear three evidentiary hurdles: it needs high-quality input data, rigorous ground-truthing across ecoregions, and independent validation against established inventories. The Microsoft award explicitly targets the first two—new lidar flights, expanded field sampling, and iterative model refinement with university partners. But publicly available peer-reviewed validation of TealWaters’ exact implementations remains thin. Until state agencies or third-party researchers publish accuracy metrics across seasons and habitat types, performance claims deserve cautious optimism.
The potential upside is enormous. Pinpointing carbon-rich peatlands lets governments target protection to prevent massive greenhouse gas releases. Municipal planners gain low-cost, nature-based flood-mitigation blueprints at a time when extreme precipitation events are intensifying. Small counties can sidestep expensive consultant studies by tapping open or low-cost high-resolution layers. Conservation groups gain a prioritization engine that triages projects where every dollar yields the highest return in biodiversity, water quality, or carbon sequestration.
Those benefits, however, travel with a set of thorny ethical and practical risks. Historical wetland loss data and ground-truth observations are heavily skewed toward temperate regions; a model trained on Washington’s ecosystems may fail dramatically in the tropics or the Arctic, where peatlands store immense carbon but lack decades of field notes. A bright green pixel on a WIP map is a probability signal, not a legal boundary—misapplied by a planning commission, it could block legitimate development or lull a community into a false sense of security. Tribal partners, including the Tulalip, rightly insist that culturally sensitive geospatial data remain under their governance, not siphoned into a commercial cloud without consent. The deal also raises the specter of “enabled emissions”—the downstream greenhouse gases that cloud-based AI makes possible when its outputs are co-opted by extractive industries. Without explicit usage policies, detailed maps of high-carbon wetlands could become a roadmap for resource exploitation rather than protection.
From an engineering perspective, the challenges are equally formidable. Forested wetlands hide beneath canopy that satellite optical sensors cannot penetrate; lidar and terrain-wetness indices help but depend on data collection that many jurisdictions cannot afford. Wetlands are seasonally dynamic, so models must ingest time-series radar and multiyear imagery and output probabilistic surfaces that reflect temporal variability. Planners use a hodgepodge of GIS platforms, demanding outputs in standard formats like GeoJSON and OGC web services, meticulously documented so they slot into existing workflows. Carbon accounting at landscape scale requires field cores and ecosystem-specific allometric models; without them, carbon-stock maps carry confidence intervals too wide to support carbon-market verification.
To steer this toward genuine societal value, governments and funders should enact a handful of practical safeguards. Every grant recipient should be required to publish confusion matrices, regional accuracy breakdowns, and independent validation reports. Field-sampling campaigns in underrepresented ecoregions must be funded to close data gaps. Data-sovereignty clauses need to be embedded in contracts, giving tribal and community partners permanent control over culturally sensitive layers and a veto on public release. Outputs should include per-pixel uncertainty bands and recommended policy thresholds, so that a county planner knows exactly when a WIP score is high enough to trigger a permit review. And cloud providers should be compelled to disclose the compute carbon footprint of these AI models and outline mitigation strategies.
Early success will be measured not in press releases but in concrete metrics: mapped wetland acreage newly discovered compared with legacy inventories, confirmed by boots-on-the-ground verification; documented instances where a WIP layer changed a permitting decision or a tribal restoration plan; publicly released model-performance dashboards that let outside researchers reproduce the results; and integration of the data into benefit-cost analyses for flood-mitigation bonds or carbon-credit registries.
Microsoft’s TealWaters bet is a high-profile test of whether cloud-scale AI can shift the trajectory of wetland conservation. The scientific scaffolding—from the 2023 Nature wetland-loss synthesis to the WIP methodology—is sound. But the gap between a clever algorithm and durable conservation is wide, bridged only by transparent validation, community governance, and policy frameworks that harness the maps for restoration rather than extraction. Washington is the proving ground; what happens there will signal whether the initiative scales beyond a proof-of-concept into a tool that genuinely redraws the conservation map.