DeepMind and WRI map global forest types at ten-metre resolution

Google DeepMind and the World Resources Institute have released ForTy, a global open-access forest mapping dataset that classifies tree cover at a ten-metre spatial resolution.

The dataset categorises global forest canopy into six functional classifications: dry forests, moist forests, temperate or boreal needleleaf, temperate or boreal broadleaf, mangroves, and planted forests. Built on DeepMind's AlphaEarth Foundations model, the system achieves an overall classification accuracy of 90.2 per cent against ground-truth validation datasets.

Previous global land-cover maps have routinely struggled to distinguish commercial tree plantations from primary, biodiverse woodlands. Conventional remote-sensing approaches relied on normalised difference vegetation indices and static spectral thresholds that often registered dense timber monocultures as intact natural forest. ForTy addresses this limitation by explicitly isolating planted stands from indigenous ecosystems, providing a substantially clearer baseline for carbon accounting and conservation planning.

Under the hood, the system leverages AlphaEarth Foundations, DeepMind's geospatial model trained on multimodal satellite feeds and environmental data. DeepMind designed the underlying architecture to act as a generalist backbone for Earth observation, transforming raw satellite data streams into unified spatial embeddings. The pipeline ingests optical observations from the European Space Agency's Sentinel-2 constellation alongside synthetic aperture radar data, topographical contours, and bioclimatic variables.

By learning contextual representations across spatial, temporal, and spectral dimensions, the architecture resolves subtle structural differences between natural, uneven canopies and systematically spaced commercial plantations. Downstream classification heads then fine-tune on these embeddings to map fragmented landscapes that simpler pixel-level classifiers mischaracterise.

The release arrives as environmental bodies and international supply chains face tighter regulatory mandates, including the European Union Deforestation Regulation. Accurate parcel-level distinctions between natural ecosystems and commercial timber dictate whether land-use transitions comply with legally binding sustainability standards.

DeepMind and the World Resources Institute have published the dataset across Google Earth Engine and open-data platforms. Whether foundation models trained on recent satellite archives can maintain this level of classification fidelity as climate extremes alter baseline phenology remains an open technical challenge.