Why understanding the drivers of forest loss matters
Ongoing forest loss and degradation due to human activities, including production and extraction of commodities, continue to be an urgent challenge, leading to biodiversity loss, carbon emissions, and negative impacts on communities. Recognising this challenge, governments and companies have undertaken various commitments and efforts to mitigate negative impacts on forests, including supply chain-focused initiatives like voluntary corporate zero-deforestation commitments and demand-side regulations such as the EU Deforestation Regulation (EUDR).
Over the past decade, advances in remote sensing have improved our ability to monitor global forest dynamics, including annual and near real time disturbances, enabling greater transparency, accountability, and action. However, disturbances to tree cover can occur due to a variety of reasons—including natural ecological cycles (like windthrow or river meandering); management (such as harvest cycles in wood fiber plantations); or land use changes that are considered permanent (when tree cover is cleared for cropland and pastures, mining, or built infrastructure). Without knowing the driver of loss, it can be challenging to pinpoint when and where tree cover loss is a cause for concern, as well as understand the potential impacts of these disturbances. Our research—a collaboration between World Resources Institute and Google DeepMind— aims to fill this gap in order to enable targeted solutions to protect and sustainably manage the world’s forests.
Our approach to mapping drivers: how the data was created
To better understand what drives global forest loss and inform effective interventions, our study set out to map the dominant causes of tree cover loss across the world. We used a deep learning model based on satellite imagery and other ancillary datasets to produce a global map of the dominant drivers of tree cover loss[1] from 2001-2024[2] at 1 km spatial resolution, making it possible to identify the cause of loss at a more detailed spatial scale than was possible previously with available global datasets.
In our study, we map the dominant driver, which refers to the driver that caused the majority of tree cover loss (as detected by the University of Maryland’s tree cover loss product available on Global Forest Watch) from 2001-2024 within each 1 km cell. We classify seven different driver categories: permanent agriculture (including seasonal & perennial crops, tree crops, and pasture), hard commodities (mining and energy infrastructure), logging, shifting cultivation[3], wildfire, settlements and infrastructure, and other natural disturbances.
To train the model, we created a training dataset through visual interpretation of nearly 7,000 samples around the world using very-high resolution imagery over time to label the driver of tree cover loss. We developed a customised Residual Network (ResNet), a type of deep learning model, which uses the labeled training samples to learn how to associate the patterns from the input data with a driver class. Our accuracy assessment of the final map using an independent stratified sample resulted in an overall accuracy of 90.5%.
What do the results show?
Our results show that permanent agriculture is the leading driver of tree cover loss globally, representing 33% of all tree cover loss from 2001-2024, or 168 million hectares—an area larger than the size of Mongolia. Permanent agriculture was the predominant driver of loss across the tropics. Wildfire was the second largest driver of tree cover loss globally, representing 29% of all tree cover loss (151 million hectares), followed by logging, representing 25% (131 million hectares). While other drivers make up a smaller proportion of tree cover loss globally, they may represent important regional or local drivers (Figure 2).
Figure 1: Drivers of tree cover loss by region. The area displayed within the circle for each region represents total tree cover loss from 2001-2024, with the proportion attributed to each driver represented by color.
Figure 2: Examples of each driver class. Color opacity corresponds to tree cover loss intensity within each 1 km grid cell. The Global Forest Change tree cover loss data at 30m resolution is displayed in white beneath the drivers of tree cover loss map. From left to right: 1) agricultural expansion in Bolivia, 2) gold mining in Peru, 3) shifting cultivation in Democratic Republic of Congo, 4) routine timber harvest in managed forests in Sweden, 5) wildfire in Russia, 6) urban expansion in China, 7) tree mortality due to bark beetle in Colorado, USA.
The data allows us to separate losses that are more likely to be temporary versus those that are likely to be permanent. Although our analysis of the data does not take tree cover gain into account and the data does not monitor regrowth, drivers like permanent agriculture, hard commodities, and settlements and infrastructure are more likely to represent a permanent land use change and comprised 34% of all tree cover loss from 2001-2024. Drivers like logging, wildfire, shifting cultivation, and other natural disturbances comprised 66% of all tree cover loss and are more likely to represent a temporary disturbance that may be followed by regrowth. Permanent land use change, or deforestation, often has more severe impacts compared to temporary disturbances, such as permanent loss of carbon stocks and more profound habitat disruption. However, this doesn’t mean these temporary disturbances won’t impact forests, since forests can take decades to regrow and their condition after regeneration can vary.
How can the data be used to inform supply chain sustainability initiatives?
Geospatial monitoring plays a critical role in advancing sustainable supply chains by providing information that can be used to track progress, monitor and mitigate threats, and develop effective strategies—particularly as companies continue to improve supply chain traceability. Stakeholders from different sectors can use the insights from this data in a variety of ways:
Civil society and researchers can use the data to assess progress on forest related targets and hold stakeholders accountable. Spatially detailed and globally consistent data is important to assess progress across regions and ensure comparisons between geographies are based on consistent methods and definitions. This data can help provide a comprehensive view on global progress toward ending agriculture-driven loss that is often the focus of supply chain initiatives.
Companies can use this data as a source of useful information to support monitoring and identification of risks within supply chains by helping to identify where tree cover is being cleared for commodities, including agricultural products, hard commodities like minerals, oil and gas, or timber.
However, within this context, some cautions should be taken into account. Following the tree cover loss data set, we adopt a broad definition of tree cover which can include both natural forests or planted trees: woody vegetation greater than 5 meters in height. Supply chain initiatives are often focused on monitoring forests as defined using a more limited set of criteria (minimum area threshold or land use designation), or natural forests specifically. To align with stricter definitions, additional data from other sources can be combined with our data on drivers to provide a baseline for analysis and exclude any areas that don’t qualify as forests.
Policymakers and land managers can use this data to support jurisdictional monitoring by identifying the causes of forest disturbances within their jurisdictions and determining the most appropriate interventions depending on the local context. As mentioned above, additional data can be combined with our drivers data to hone in on specific areas of interest. For example, data on primary forests or protected areas can be combined with our drivers data to help separate routine harvesting in wood fiber plantations from illegal logging in primary forests—helping to characterise where logging might pose a threat to forest health and develop strategies accordingly.
For more information on the data and how it was produced, visit the publication, and read more from WRI.
[1] In this study, we define ‘driver’ as the direct cause of disturbances to tree cover. ‘Tree cover loss’ is defined, following Hansen et al. 2013, as a stand-replacement disturbance to woody vegetation greater than 5 meters in height.
[2] The Environmental Research Letters publication covers the period from 2001-2022. We have since updated the data for 2001-2024 and the results presented in this blog represent the updated data. Statistics in this blog were calculating using a 30% canopy density threshold.
[3] Shifting cultivation is a type of rotational, subsistence agriculture traditionally practiced in the tropics where forests are temporarily cleared to cultivate crops, and then left fallow for many years to allow forests to regrow. In our study this is distinguished from permanent agriculture, where agricultural activity persists for many years and is not part of a temporary cultivation cycle.
