source
Bart Slagter, Johannes Reiche, Diego Marcos, Adugna Mullissa, Etse Lossou, Marielos Peña-Claros, Martin Herold, Monitoring direct drivers of small-scale tropical forest disturbance in near real-time with Sentinel-1 and -2 data, Remote Sensing of Environment, Volume 295, 2023, 113655, ISSN 0034-4257, https://doi.org/10.1016/j.rse.2023.113655.(https://www.sciencedirect.com/science/article/pii/S0034425723002067)
overview
This dataset enhances pantropical near-real-time deforestation monitoring systems by using a deep learning model to classify direct drivers of forest disturbances. Produced by Wageningen University and Research (WUR), Laboratory of Geo-information Science and Remote Sensing, with support from Global Forest Watch (GFW), the dataset is built upon GFW's integrated deforestation 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. 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.
**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 Amazon forest (Bolivia, Brazil, Colombia, Ecuador, French Guiana, Guyana, Peru, Suriname and Venezuela), Congo Basin forest (Cameroon, Central African Republic, Democratic Republic of the Congo, Equatorial Guinea, Gabon, Republic of the Congo) and Southeast Asian forest (Brunei, East- Timor, Indonesia, Malaysia and Papua New Guinea).
The model was trained on a dataset of labeled deforestation 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 displays historical alert data (2022-2024), the classifications were applied in a simulated near-real-time monitoring scenario to classify monthly deforestation alerts into one of the specified classes, achieving an overall accuracy of 0.89 Macro-F1. Historical data were classified up to three times within the three months following each alert, with the most confident result retained. Classifications are only reported when at least five connected pixels share the same class at a confidence score above 0.75; otherwise, alerts remain unlabeled. Accuracy varies by region, with the Congo Basin performing best. Most misclassifications stem from confusion between agriculture and mining, which often co-occur, and between flooding and wildfires.
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 low post-disturbance normalized-burn ration 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, 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 deforestation alerts in the pantropical region will be available. The driver attributions will be continuously mapped and openly distributed via Global Forest Watch, as an extension to the integrated deforestation alerts.
See the integrated deforestation alert layer [metadata](https://www.globalforestwatch.org/map/?map=eyJkYXRhc2V0cyI6W3siZGF0YXNldCI6ImludGVncmF0ZWQtZGVmb3Jlc3RhdGlvbi1hbGVydHMtOGJpdCIsIm9wYWNpdHkiOjEsInZpc2liaWxpdHkiOnRydWUsImxheWVycyI6WyJpbnRlZ3JhdGVkLWRlZm9yZXN0YXRpb24tYWxlcnRzLThiaXQiXX0seyJkYXRhc2V0IjoicG9saXRpY2FsLWJvdW5kYXJpZXMiLCJsYXllcnMiOlsiZGlzcHV0ZWQtcG9saXRpY2FsLWJvdW5kYXJpZXMiLCJwb2xpdGljYWwtYm91bmRhcmllcyJdLCJvcGFjaXR5IjoxLCJ2aXNpYmlsaXR5Ijp0cnVlfV19&modalMeta=gfw_integrated_alerts) for additional details about the deforestation alerts.
The alert drivers are available on Google Earth Engine with asset IDs:
- projects/wurnrt-drivers/assets/distribution/driverclassification_afr_202201_202412
- projects/wurnrt-drivers/assets/distribution/driverclassification_sa_202201_202412
- projects/wurnrt-drivers/assets/distribution/driverclassification_sea_202201_202412
cautions
- This dataset focuses on classifying the key drivers of forest disturbances and may not include all potential causes of deforestation.
- In the future, this dataset will be updated on a monthly basis. The model will reclassify all alerts detected in the most recent 3 months, so alerts may be updated up to 3 times. Since the most confident classification will be shown, classifications will be subject to change for up to 4 months after the initial alert.
- Classifications are only reported when at least 5 connected pixels share the same driver at a confidence score of 0.75 or higher.
- Alerts may remain unlabeled if no driver reaches the 0.75 confidence threshold or if post-disturbance land-use signals do not appear in imagery within 3–4 months. In the historical dataset (2022–2024), only 4.3% of alerts were left unclassified.
- The term deforestation is used because these are *potential* deforestation events, and alerts could be further investigated to determine this.
- We do not recommend using deforestation alerts for global or regional trend assessment, nor for area estimates. Rather, we recommend using the annual tree cover loss data for a more accurate comparison of the trends in forest change over time, and for area estimates. Recent alerts will include false positives that have yet to raise their confidence level and may eventually be removed. Past alerts may have been removed in error from the database if rapid canopy closure precedes the additional unobscured satellite observations within 6 months. Additionally, updates to the methodologies and variation in cloud cover between months and years pose additional risks to using deforestation alerts for inter/intra-annual comparison.
- When zoomed out, this data layer displays some degree of inaccuracy because the data points must be collapsed to be visible on a larger scale. Zoom in for greater detail.