Detecting forest disturbances in Germany from satellite time series using unsupervised LSTM autoencoders

Climate change and natural disturbances are posing an increasing pressure on European forests and have led to extensive forest losses over recent decades. Large-scale monitoring of forest dynamics is therefore essential and can be effectively supported by remote sensing techniques. Previous studies have demonstrated the potential of supervised and unsupervised machine learning approaches for detecting forest disturbances, typically characterising forest condition at specific points in time. At the same time, comprehensive reference data remain scarce at large spatial scales.

Against this background, we investigate the potential of an unsupervised deep learning approach for large-scale detection of forest anomalies across Germany within the framework of the EO4Nature project. We apply a deep learning based Long Short-Term Memory (LSTM) autoencoder to model vegetation trajectories over multiple vegetation periods to capture gradual changes in forest vitality.

The LSTM model is trained on stratified healthy forest pixels across Germany, selected based on a low disturbance probability derived from the European Forest Disturbance Atlas (EFDA). We compare multiple model configurations using different input feature sets based on Sentinel-2 data at a monthly temporal resolution for the period 2018-2025. Anomalies in forest vitality are detected based on the reconstruction error of the autoencoder, using adaptive thresholds that account for seasonal variation and forest type. This enables the identification of pixels with different levels of anomaly severity.

We primarily evaluate the proposed approach using independent disturbance reference data at the local scale. High-resolution annual orthophotos from multiple disturbed forest sites in Germany are used to enable a detailed spatial assessment of detected anomalies.

In addition, we conducted a preliminary large-scale consistency check by comparing areas exhibiting high anomaly scores with disturbed forest regions derived from the EFDA. These initial results indicate that the unsupervised LSTM autoencoder, trained on stable forest conditions using NDVI, NBR and abiotic variables, produces a continuous anomaly score that correlates with independently mapped disturbance patterns (Spearman’s ρ = 0.65, p < 0.001), demonstrating consistency with external disturbance probabilities.

The results give insight into the disturbance intensities at which deviations from healthy forests dynamics become detectable and provide knowledge about the most relevant spectral features for large-scale monitoring of forest ecosystem stability.