The World Resources Institute and Google DeepMind created a global map of the dominant drivers of tree cover loss from 2001 to 2023 at 1 km spatial resolution (v1.1). We used a deep learning model to classify seven driver classes: permanent agriculture, hard commodities, shifting cultivation, logging, wildfires, settlements & infrastructure, and other natural disturbances. As part of the study, we collected a set of training samples through interpretation of very high resolution satellite imagery and developed a single world-wide customized residual convolutional neural network model (ResNet) using satellite data (Landsat and Sentinel-2) and ancillary biophysical and population data. In addition, we collected a stratified random sample of validation plots through interpretation of very high resolution satellite imagery to estimate the accuracy of the final classification map. For a full description of the methods, technical specifications, definitions, and limitations, please see the publication: https://doi.org/10.1088/1748-9326/add606. In this repository, we make the following available: The training and validation data, available in two separate files. Both datasets were collected by a team of image interpreters and assessed for quality by two additional interpreters. Note that while for the validation data the quality of the primary driver was rigorously assessed, the secondary driver wasn’t subject to the same level of quality control. The global drivers of forest loss raster (drivers_forest_loss_1km.tif), including the discrete classification and probabilities for each class. Creator & Contact Created by World Resources Institute and Google DeepMind Contact: Michelle Sims: Michelle.Sims@wri.org Annual updates This dataset is updated annually with the Global Forest Change product from the University of Maryland (Hansen et al., 2013). With each update, any training data plots containing the latest year of tree cover loss as the mode loss year are reinterpreted by image interpreters and the dominant driver label is updated, where necessary. An additional random sample of ~200 plots containing the latest year of tree cover loss is interpreted and added to the sample. Input data is updated for the Global Forest Change product, tree cover loss due to fires product (Tyukavina et al. 2022), Sentinel 2, and Dynamic World (Brown et al. 2022) and the model is retrained with the updated training data and rerun to produce a new global map representative of the new time period. The files follow this specific naming convention: Data type: `training` or `validation`; or `drivers_forest_loss_1km` for global raster Start year: `2001` End year: `2022` or `2023` or `2024` Version: `v1` or `v1_1` or `v1_2` Definitions A driver is defined as the direct cause of tree cover loss, and can include both temporary disturbances (natural or anthropogenic) or permanent loss of tree cover due to a change to a non-forest land use (e.g., deforestation). The dominant driver is defined as the direct driver that caused the majority of tree cover loss within each 1 km cell over the time period. Classes are defined as follows: Driver Definition Permanent agriculture Long-term, permanent tree cover loss for small- to large-scale agriculture. This includes perennial tree crops such as oil palm, cacao, orchards, nut trees, and rubber, as well as pasture and seasonal crops and cropping systems, which may include a fallow period. Agricultural activities are considered "permanent" if there is visible evidence that they persist following the tree cover loss event and are not a part of a temporary cultivation cycle. Clearing land for agricultural activities may involve use of fire. Hard commodities Tree cover loss due to the establishment or expansion of mining or energy infrastructure. Mining activities range from small-scale and artisanal mining to large-scale mining. Energy infrastructure includes power lines, power plants, oil drilling and refineries, wind and solar farms, flooding due to the construction of hydroelectric dams, and other types of energy infrastructure. Shifting cultivation Tree cover loss due to small- to medium-scale clearing for temporary cultivation that is later abandoned and followed by subsequent regrowth of secondary forest or vegetation. Clearing land for temporary cultivation may involve use of fire. Logging Forest management and logging activities occurring within managed, natural or semi-natural forests and plantations, often with evidence of forest regrowth or planting in subsequent years. This includes harvesting in wood-fiber plantations, clear-cut and selective logging, establishment of logging roads, and other forest management activities such as forest thinning and salvage or sanitation logging. Wildfire Tree cover loss due to fire with no visible human conversion or agricultural activity afterward. Fires may be started by natural causes (e.g. lightning) or may be related to human activities (accidental or deliberate). Settlements and infrastructure Tree cover loss due to expansion and intensification of roads, settlements, urban areas, or built infrastructure (not associated with other classes). Other natural disturbances Tree cover loss due to other non-fire natural disturbances, including storms, flooding, landslides, drought, windthrow, lava flows, sediment flow or meandering rivers, natural flooding, insect outbreaks, etc. If tree cover loss due to natural causes is followed by salvage or sanitation logging, it is classified as logging.

The training and validation data have the following columns for each plot: Column name Description ID Unique ID for each plot. Latitude Latitude for center point of 1 km plot. Longitude Longitude for center point of 1 km plot. Driver_primary_code Numeric code assigned to each driver: Permanent agriculture (1), Hard commodities (2), Shifting cultivation (3), Logging (4), Wildfires (5), Settlements & infrastructure (6), Other natural disturbances (7). *Noise/non-forest (8) *Note– this class is only included in the training data, in place of a separate “Noise” column. See below. Driver_primary Primary driver accounts for >50% of all loss area within the study period in 1 km plot. Confidence_primary Image analyst's confidence in interpretation of primary driver: low, medium or high. Driver_secondary Secondary driver accounts for <50% of all loss area within the study period in 1 km plot. Secondary driver includes a category called Commodity-driven deforestation which includes both permanent agriculture and hard commodities. Most but not all notes fields include the type of commodity-driven deforestation as secondary driver. Confidence_secondary Image analyst's confidence in interpretation of secondary driver: low, medium or high. Active_learning* Binary variable indicating 1 if the plot was collected as part of additional data collection effort via active learning. *Note: this column is only included in the training data. Noise* True/False field with True label assigned to tree cover loss pixels that are noise (e.g. due to cloud/sensor issues, shadows, etc.) or that never had tree cover or shrubby vegetation throughout the time series. *Note: this column is only included in the validation data. Region Regions are North America, Latin America, Europe, Africa, and Oceania. We split Asia into two subregions: 1) Southeast Asia and the rest (East/South/Central Asia & MiddleEast), and we refer to these areas as Southeast Asia and Asia. Region_code Numeric code assigned to each region: North America (1), Latin America (2), Europe (3), Africa (4), Asia (5), Southeast Asia (6), and Oceania (7). Tags Tags are a limited set of labels that provide further information on the category of commodity-driven deforestation in the 1km plot: Mining, Energy, Agriculture, Tree crop management. Tags can be associated with either Driver_primary if Driver_primary_code is 1 or 2, or it can be associated with Driver_secondary if Driver_primary_code is not 1 or 2. Notes Optional field that provides contextual information on the driver in the 1 km plot, or provides information on why a medium or low confidence label was assigned.

The global drivers of forest loss raster has the following specifications: Resolution: 0.01 degrees (3/5 arcsec, approximately 1 km at equator) Spatial extent: global, where forest loss was detected according to the Global Forest Change product [Hansen et al. 2013] Temporal extent: aggregated 23 years (2001-2023) Bands (uint8 data type): Band 1: Most likely class based on raw probabilities (values from 1 to 7). Band 2: Probability of "Permanent agriculture" class (scaled to [0-250]). Band 3: Probability of "Hard commodities" class (scaled to [0-250]). Band 4: Probability of "Shifting cultivation" class (scaled to [0-250]). Band 5: Probability of "Logging" class (scaled to [0-250]). Band 6: Probability of "Wildfire" class (scaled to [0-250]). Band 7: Probability of "Settlements and infrastructure" class (scaled to [0-250]). Band 8: Probability of "Other natural disturbances" class (scaled to [0-250]).