Time Series Analysis of Sentinel-2 Imagery for Mapping Forest Disturbance Agents

Forests play a vital role in regulating the Earth's climate as they are the largest terrestrial carbon sinks. Climate change is increasing the degradation of trees throughout Europe due to disturbance from biotic agents such as insect outbreaks (e.g. bark beetle), abiotic factors such as drought and windthrow, wildfires, and anthropogenic impacts such as logging. Dense satellite imagery provides an opportunity to accurately detect disturbance and determine the timing that disturbances occurred, but determining the driving force behind these disturbances continues to be a challenge.

Recent time series analysis (TSA) methods, particularly the Forest Disturbance Level (FDL) framework, have shown strong capability in detecting forest disturbances using Sentinel-2 imagery. By modeling forest phenology and cumulative anomaly patterns, FDL-derived metrics, such as the Forest Disturbance Date (FDD) and cumulative deviation measures, provide detailed information on the timing, duration, and severity of disturbances. However, this method cannot identify disturbance agents.

This research proposes to expand the use of the FDL framework from the detection of disturbances to the identification of disturbance agents. The proposed method enhances the FDL model by incorporating detailed phenological modeling and data‐driven feature selection. The extraction of spectral bands and vegetation indices is performed first using Sentinel-2 time series data. Next, a combined correlation analysis and Random Forest-based feature importance ranking is conducted to identify the most informative spectral bands and vegetation indices. The proposed approach uses TSA-based breakpoint detection methods. This combined framework incorporates temporal descriptors of disturbance and applies machine-learning techniques after a disturbance has been detected. After disturbances have been detected, new variables can be calculated to describe post-disturbance behavior. Based on these variables, their potential for discriminating between disturbance drivers is analyzed using Random Forest classifiers. Variables developed through the use of FDL time series analyses can also be used to describe recovery dynamics after a disturbance and disturbance trend behavior. They can additionally characterize phenological shifts and spectral patterns associated with the disturbance events of interest.

This framework is applied to Sentinel-2 surface reflectance time series spanning 2020–2024 across European forests, using reference data from the European Forest Disturbance Atlas and Copernicus forest type maps. Preliminary results suggest that post-disturbance temporal and phenological features capture informative patterns associated with different disturbance processes.

This study seeks to improve the field of forest monitoring from simply identifying disturbances to analyzing the possible attribution of disturbance types by utilizing a combination of variables based on the analysis of Sentinel-2 time series data. This study also aims to create a foundation for future analyses to identify the possible drivers of forest degradation and the factors of forest vulnerability at a larger scale.