A new study in Remote Sensing of Environment introduces a deep learning approach that maps forest disturbances across continental Europe, going back to 1985. The work forms the basis of version 3 of the European Forest Disturbance Atlas (EFDA), the longest continental-scale forest disturbance record available from satellite data. 

EFDA was developed by researchers at the Technical University of Munich as part of the ForestPaths project. The atlas combines the Landsat satellite archive with machine learning to produce standardised, annual disturbance data across 38 European countries. 

The new study replaces the atlas's original Random Forest-based approach with two deep learning architectures trained on nearly 40 years of satellite imagery. The better-performing model, a 1D U-Net, consistently outperformed both a second deep learning model and the previous Random Forest method. 

The result is a new annual, 30-metre-resolution map of forest disturbance across Europe, covering 216 million hectares of forest and revealing a total disturbed area of 48.5 million hectares since 1985. This represents over a fifth of Europe's entire forest cover. Compared to the previous version of EFDA, the latest update achieves a better balance between two types of error: missing real disturbances versus flagging false ones. This has so far been a persistent challenge for automated forest monitoring at this scale. The model performed especially well in complex landscapes that have tripped up older methods. 

Because the model needs only short sequences of yearly data to make predictions, the researchers designed it to be easily updated as new satellite images become available. 

Read the full study.