DeepMind details ForTy to map global forest types at 10-metre resolution

Google DeepMind has detailed ForTy, a remote-sensing model that maps global forest varieties at a 10-metre spatial resolution.

Existing Earth observation models routinely estimate canopy height, tree cover, and aggregate biomass. However, these systems often fail to distinguish between ecologically distinct forest structures. Satellite datasets have historically operated at coarse resolutions between 100 metres and one kilometre, or collapsed diverse biomes into generic vegetation classes. Consequently, automated audits frequently conflate biodiverse primary woodlands with commercial monoculture plantations.

ForTy targets these classification bottlenecks by training on 1.7 million ground-truth field locations worldwide. DeepMind sourced the dataset from distributed forest inventories and regional ecological records. By pairing verified ground plots with European Space Agency Sentinel-2 multispectral imagery, the model learns spectral signatures tied to distinct canopy architectures and botanical compositions.

Operating at 10-metre resolution allows the model to resolve fine-grained habitat fragmentation, ecotones, and localised logging activity. The classification schema covers specific forest types. This allows users to evaluate biodiversity quality, audit restorative forestry, and calibrate carbon accounting pipelines. Accurate species-group mapping remains critical for carbon estimation, as wood density and carbon sequestration rates vary substantially between tree varieties.

DeepMind developed ForTy to supply a uniform, globally consistent baseline for environmental monitoring. Prior mapping efforts often relied on fragmented, country-specific methodologies, producing incompatible datasets that complicated cross-border conservation. By grounding satellite predictions in standardised field data, the system attempts to bridge the gap between local forestry inventories and orbital observation.

The project highlights an ongoing pivot within remote sensing toward granular ecological functional types rather than surface-level cover metrics. How readily conservation agencies, carbon registries, and regional governments integrate these 10-metre classifications into operational policy remains an open question.