Deforestation alerts (RADD)

This dataset is a forest loss alert product developed by the GLAD (Global Land Analysis and Discovery) lab at the University of Maryland. GLAD-S2 alerts utilize data from the European Space Agency’s Sentinel-2 mission, which provides optical imagery at a 10 m spatial resolution with a 5-day revisit time. The shorter revisit time, when compared to GLAD Landsat alerts, reduces the time to detect forest loss and between the initial detection of forest loss and classification as high confidence. This is particularly advantageous in wet and tropical regions, where persistent cloud cover may delay detections for weeks to months. GLAD-S2 alerts are available for primary forests in the Amazon basin from January 1st, 2019, to present, updated daily.  New Sentinel-2 images are analyzed as soon as they are acquired. Cloud, shadow, and water are filtered out of each new image, and a forest loss algorithm is applied to all remaining clear land observations. The algorithm relies on the spectral data in each new image in combination with spectral metrics from a baseline period of the previous two years.  Alerts become high confidence when at least two of four subsequent observations are flagged as forest loss (this corresponds to “high,” “medium,” and “low” confidence loss on the GLAD app linked below). The alert date represents the date of forest loss detection. Users can choose to display only high confidence alerts on the map, but keep in mind this will filter out the most recent detections of forest loss. Additionally, forest loss will not be detected again on pixels with high confidence alerts. Alerts that have not become high confidence within 180 days are removed from the dataset.  The GLAD-S2 alerts are available on \*\*Google Earth Engine\*\* with asset ID: projects/glad/S2alert  Monitor primary forest loss in near-real time using Sentinel-2 imagery

This dataset, assembled by Global Nature Watch, aggregates deforestation alerts from four alert systems (DIST-ALERT, GLAD-L, GLAD-S2, RADD) into a single, integrated disturbance alert layer. This integration allows users to detect deforestation and other vegetation loss faster than any single system alone, as the integrated layer is updated when any of the source alert systems are updated.   The source alert systems are derived from satellites of varying spectral and spatial resolutions. The 30 m GLAD Landsat-based alerts and DIST-ALERTs are up-sampled to match the 10 m spatial resolution of purely Sentinel-based alerts (GLAD-S2, RADD). Using the same 10 m grid avoids the double counting of overlapping alerts, which are instead classified at a higher confidence level, indicated by darker pixels.   Alerts are classified as high confidence when detected multiple times (two or more times for GLAD-L, GLAD-S2 and RADD, and four or more times for DIST-ALERT) by a single alert system. This can occur in areas and at times when only one alert system was operating. Where multiple alert systems are operating, alerts detected by multiple (two or more) of these systems are classified as highest confidence. With multiple sensors picking up change in the same location, we can be more confident that an alert was not a false positive and do not need to wait for additional satellite imagery to increase confidence in detected loss, thus providing more confident alerting faster than with a single system. Alerts from multiple systems which are within 180 days (6 months) of each other are considered a single event. In this case, the earliest date within that 180 day window will appear as the alert date in the integrated alert product. These will be considered highest confidence alerts. The date of the alert remains until another new alert (which is not within 6 months) in the same pixel is registered.   A study conducted by Wageningen University in collaboration with researchers from Global Nature Watch and University of Maryland's GLAD lab found that integrating alert systems results in faster detection of new disturbances by days to months, and also shortens the delay to increase confidence. Combined alerts have a higher producer's accuracy (fewer false negatives), but a lower user's accuracy (more false positives) since the commission errors from each system are combined; however, 'highest confidence' alerts, where more than one system detected the change, effectively eliminated false detections. Learn more: https://iopscience.iop.org/article/10.1088/1748-9326/ad2d82   Monitor vegetation disturbance in near-real-time using integrated alerts from multiple alerting systems

This dataset, assembled by Global Nature Watch, aggregates deforestation alerts from three alert systems (GLAD-L, GLAD-S2, RADD) into a single, integrated deforestation alert layer. This integration allows users to detect deforestation events faster than any single system alone, as the integrated layer is updated when any of the source alert systems are updated. The source alert systems are derived from satellites of varying spectral and spatial resolutions. 30 m GLAD Landsat-based alerts are up-sampled to match the 10 m spatial resolution of Sentinel-based alerts (GLAD-S2, RADD). This avoids the double counting of overlapping alerts, which are instead classified at a higher confidence level, indicated by darker pixels. Alerts are classified as high confidence when detected twice by a single alert system. This can occur in areas and at times when only one alert system was operating. Where multiple alert systems are operating, alerts detected by multiple (two or three) of these systems are classified as highest confidence. With multiple sensors picking up change in the same location, we can be more confident that an alert was not a false positive and do not need to wait for additional satellite imagery to increase confidence in detected loss, thus providing more confident alerting faster than with a single system. A study conducted by Wageningen University in collaboration with researchers from Global Nature Watch and University of Maryland's GLAD lab found that integrating alert systems results in faster detection of new disturbances by days to months, and also shortens the delay to increase confidence. Combined alerts have a higher producer's accuracy (fewer false negatives), but a lower user's accuracy (more false positives) since the commission errors from each system are combined; however, 'highest confidence' alerts, where more than one system detected the change, effectively eliminated false detections. Learn more: https://iopscience.iop.org/article/10.1088/1748-9326/ad2d82 The integrated deforestation alerts are available on **Google Earth Engine** with asset ID: projects/forma-250/assets/gfw_integrated_alerts/default_latest **Download** the raster tiles here: https://data.globalforestwatch.org/datasets/gfw::integrated-deforestation-alerts/about Monitor forest disturbance in near-real-time using integrated alerts from three alerting systems

This dataset, created by the [[GLAD]](http://glad.geog.umd.edu/) (https://glad.geog.umd.edu/) (Global Land Analysis & Discovery) lab at the University of Maryland and supported by Global Nature Watch, is the first Landsat-based alert system for tree cover loss. While most existing loss alert products use 250-meter resolution MODIS imagery, these alerts have a 30-meter resolution and thus can detect loss at a much finer spatial scale. These alerts are operational for land areas between 30 degrees north and south.  New Landsat 8 and 9 images are downloaded as they are posted online, assessed for cloud cover or poor data quality, and compared to the three previous years of Landsat-derived metrics (including ranks, means, and regressions of red, infrared and shortwave bands, and ranks of NDVI, NBR, and NDWI). The metrics and the latest Landsat image are run through seven decision trees to calculate a median probability of forest disturbance. Pixels with probability >50% are reported as tree cover loss alerts. The entire process is run in Google Earth Engine to ensure reliable updates and scalability. For more information on methodology, see the paper in Environmental Research Letters.  Alerts are not classified as high confidence until two or more out of four consecutive observations are labelled as tree cover loss. Alerts are removed from the dataset after four consecutive observations or more than 180 days if they are not classified as high confidence. You can choose to view only high confidence alerts in the menu, though keep in mind that using only high confidence alerts misses the newest detections of tree cover loss.  The GLAD-L alerts are available on \*\*Google Earth Engine\*\* with asset ID: projects/glad/alert/UpdResult  Monitor tree cover disturbance in near-real-time using Landsat imagery

This dataset enhances pantropical near-real-time forest disturbance monitoring systems by using a deep learning model to classify direct drivers of forest disturbances on a monthly basis. Produced by Wageningen University and Research (WUR), Laboratory of Geo-information Science and Remote Sensing, with support from Global Nature Watch (GNW), the dataset is built upon GNW's integrated disturbance alerts, which contains WUR's RADD (Radar for Detecting Deforestation) alerts and University of Maryland's GLAD (Global Land Analysis & Discovery) alerts, harnessing post-disturbance imagery from Sentinel-1 and Sentinel-2 satellites. Only high and highest confidence alerts are classified. This approach enables the identification of these direct causes driving deforestation: small-scale agriculture, large-scale agriculture, road development, selective logging, mining, wildfire, flooding, and other natural disturbances. The alert drivers offer more targeted enforcement of laws and regulations, improved estimations of ecological impacts, and better understanding of carbon emissions related to forest disturbances. Read the paper here: Rapid satellite monitoring reveals complexity of tropical tree cover loss drivers at fine scale | Research Square **Alert Driver class definitions** **Non-natural drivers (human-caused)** **Small-scale agriculture**: Clearings smaller than 2 ha, commonly related to shifting cultivation (temporary clearing for cultivation with later regrowth) or smallholder farming. In smallholder landscapes with mixed disturbances, this may include artisanal logging and fuelwood collection (typically burned for cooking).  **Small-scale agriculture with fire**: Clearings for small-scale agriculture (see above) where fire was likely used for the clearing, observed by a coinciding VIIRS fire alert and low post-disturbance Sentinel-2 Normalized Burn Ratio **Large-scale agriculture**: Clearings larger than 2 ha for the establishment of crops or pastures, commonly related to industrial agriculture (e.g. production of soy, palm oil, beef, etc.), clearcuts, and large-scale clearings for land speculation. **Large-scale agriculture with fire**: Clearings for large-scale agriculture (see above) where fire was likely used for the clearing, observed by a coinciding VIIRS fire alert and low post-disturbance Sentinel-2 Normalized Burn Ratio **Road development**: Clearings for the establishment of roads, commonly related to facilitating industrial timber harvests, but can include roads for any purpose.  **Selective logging**: Small-scale disturbances caused by selective tree felling and skidding (paths where felled logs are dragged or transported), commonly related to industrial timber harvests.  **Mining**: Forest clearing to facilitate artisanal and industrial mineral extraction. **Natural drivers** **Flooding**: Disturbances or clearings caused by floodings and meandering rivers. This includes both natural and human-induced flooding. **Other natural disturbances**: Disturbances or clearings without visible human-induced cause. This includes windthrows, droughts, landslides and naturally dying trees. **Other drivers** **Wildfire**: Large-scale disturbances due to fire, without immediate land clearing for agricultural activity. This includes both human-induced and naturally induced fires. This class excludes controlled fires used for agricultural clearing, but includes escaped wildfires caused by controlled fires. **Unlabeled**: Pixels where a confidence threshold for a prediction is not reached. The dataset covers the primary humid tropical forest (Turubanova et al. 2018) of the Amazon (Bolivia, Brazil, Colombia, Ecuador, French Guiana, Guyana, Peru, Suriname and Venezuela, restricted to the biogeographic outlines of the Amazon biome as defined by RAISG), Congo Basin (Cameroon, Central African Republic, Democratic Republic of the Congo, Equatorial Guinea, Gabon, Republic of the Congo) and Southeast Asia (Brunei, East- Timor, Indonesia, Malaysia and Papua New Guinea). The model was trained on a dataset of labeled disturbance alerts, interpreted with high-resolution Planet imagery across the Amazon, Congo Basin, and Southeast Asia. The model integrates the complementary strengths of radar (Sentinel-1) and optical (Sentinel-2) imagery, leveraging convolutional neural networks (CNNs) to classify disturbances. Even though the dataset initially displayed historical alert data (2022-2025), the classifications were applied in a simulated near-real-time monitoring scenario to classify monthly disturbance alerts into one of the specified classes, achieving an overall accuracy of 0.89 Macro-F1. Accuracy varies by region, with the Congo Basin performing best. After the addition of the DIST-ALERT product, there are more alerts and therefore, areas of misclassification may be larger. Most misclassifications stem from confusion between agriculture and mining, which often co-occur, and between flooding and wildfires. Historical drivers (2022-2025) were classified up to three times within the three months following each alert, with the most confident result retained. These 2022-2025 classifications were postprocessed by selecting per disturbance patch the most occurring driver class, leading to some initial classifications being overwritten by another class in postprocessing. The current method, running from January 2026 onwards, selects the most occurring driver class only for disturbance patches within the three-month windows, As a result, an alert can be updated up to 3 times, and the latest classification is retained. Classifications may therefore change for up to 4 months after the initial alert and do not change after that. . A class ‘unconfident’ is built from pixels that do not exceed a 0.5 class probability from the classification output. When the most occurring class for a disturbance patch is ‘unconfident’, this alerts remains unlabeled. Post-processing steps identified where fire coincided with the clearing of agricultural land, based on the presence of VIIRS fire alerts (within a 500 m buffer around the alert and a 6-month window) and a low post-disturbance normalized-burn ratio (lower than 0.0) in the following month’s Sentinen-2 composite. Smoothing was applied to all classes except road development and selective logging, which occur at finer scales. For patches smaller than 5 km, the dominant driver was assigned. For larger patches in the 2022-2025 data, pixel-level classifications were retained but smoothed with a 5 × 5-pixel majority filter. This step was applied to the whole time series; therefore, the historical product has an advantage in differentiation between small- and large-scale agriculture compared to the operational product. In a near-real-time setting, large clearings may initially be labeled as small-scale until they exceed the 2 ha threshold. In the future, monthly driver attributions for disturbance alerts will expand to new regions. The driver attributions will be continuously mapped and openly distributed via Global Nature Watch, as an extension to the integrated disturbance alerts. We also provide an annual version of this dataset, which incorporates some post processing – specifically the differentiation between large-scale and small-scale agriculture is made at the end of the year, meaning that agricultural clearings which are identified incrementally are identified based on their size at the end of the year. See the integrated disturbance alert layer metadata for additional details about the disturbance alerts. The alert drivers are available on Google Earth Engine with asset IDs: **Latest data** (ee.ImageCollection containing data from January 1, 2026 – end of last month): - projects/wurnrt-drivers/assets/operational_postprocessed_sa_collection - projects/wurnrt-drivers/assets/operational_postprocessed_afr_collection - projects/wurnrt-drivers/assets/operational_postprocessed_sea_collection *Note: this collection updates monthly, so new images will be added over time, but the collection path itself will stay the same.* **Historical archive** (ee.Image containing data from 2022–2025): - projects/wurnrt-drivers/assets/distribution/driverclassification_afr_202201_202512 - projects/wurnrt-drivers/assets/distribution/driverclassification_sa_202201_202512 - projects/wurnrt-drivers/assets/distribution/driverclassification_sea_202201_202512 Assigns probable drivers to the integrated disturbance alerts using machine learning and Sentinel imagery

This dataset is a derivative of the OPERA’s DIST-ALERT product (OPERA Land Surface Disturbance Alert from Harmonized Landsat Sentinel-2 product), which is derived through a time-series analysis of harmonized data from the NASA/USGS Landsat and ESA Sentinel-2 satellites (known as the HLS dataset). The product identifies and continuously monitors vegetation cover change in 30-m pixels across the globe. It can be accessed through the LPDAAC website here: https://search.earthdata.nasa.gov/search?q=C2746980408-LPCLOUD, and on Google Earth Engine (GEE) with asset ID: projects/glad/HLSDIST/current The DIST-ALERT on GNW is a derivative of this, and additional data layers not used in the GNW product are available through the LPDAAC and GEE such as initial vegetation fraction, and disturbance duration. Specifically, three layers are used to create the GNW DIST-ALERT. The VEG-DIST-DATE provides the date of the alert, the VEG-DIST-COUNT is used to determine the confidence (2-3 consecutive observations for low confidence and 4+ observations for high confidence) and VEG-ANOM-MAX is used as a threshold for determining loss (>30%). While the version on the LPDAAC is updated every 2-4 days, the data is updated weekly on GNW. The product detects notable reductions in vegetation cover (measured as “vegetation fraction” or the percent of the ground that is covered by vegetation) for every pixel every time the new satellite data is acquired and the ground is not obscured by clouds or snow. The current vegetation fraction estimate is compared to the minimum fraction for the same time period (within 15 days before and after) in the previous 3 years, and if there is a reduction, then the system identifies an alert in that pixel. Anomalies of at least a 10% reduction from the minimum become alerts in the original product, and on GNW, a higher threshold of 30% is used, to reduce noise, and false alerts in the dataset. Because the product compares each pixel to the minimum for the same time period in previous years, it takes into account regular seasonal variation in vegetation cover. As the product is global and detects vegetation anomalies, much of the data may not be applicable to GNW users monitoring forests. Therefore, we mask the alerts with UMD’s tree cover map, allowing users to view only alerts within 30% canopy cover. Monitors global vegetation disturbance in near-real-time using harmonized Landsat-Sentinel-2 (HLS) imagery