ULISSE: Parameter-efficient adaptation of earth vision models to monitor forest disturbance in sentinel-2 time series

DOI: https://doi.org/10.1016/j.ecoinf.2026.103668

Forest ecosystems are increasingly vulnerable to disturbance agents under ongoing climate change. Among the most affected are coniferous trees, the dominant tree species across Europe, which have experienced severe bark beetle outbreaks over the past decade. Such disturbances are projected to intensify further as climate change progresses, with significant adverse consequences for forest health and ecosystem services. Monitoring these outbreaks therefore represents a critical ecological and forestry challenge, traditionally addressed through labor-intensive field surveys by forest rangers.

Freely available Sentinel-2 multispectral imagery, distributed under the European Copernicus programme, has recently emerged as a scalable alternative for monitoring environmental disturbances, including bark beetle-induced tree dieback. Several deep learning approaches have been proposed to map such dieback from Sentinel-2 data; however, their effectiveness is frequently constrained by the scarcity of labeled reference data required to train such classification models.

To address this limitation, we introduce ULISSE, a deep learning semantic segmentation framework for mapping forest tree dieback caused by bark beetle outbreaks using Sentinel-2 image time series. ULISSE adopts a U-Net-like architecture specifically designed to exploit multi-temporal Sentinel-2 acquisitions. The framework builds upon models initially pretrained on large volumes of Sentinel-2 imagery for land-cover classification. To adapt these pretrained models to the tree dieback mapping task, we employ Parameter-Efficient Fine-Tuning (PEFT), a parsimonious transfer learning paradigm that enables effective knowledge transfer from limited labeled training data. This strategy allows ULISSE to achieve high classification accuracy even under severe data scarcity. Experimental results demonstrate the effectiveness of the proposed approach across two case studies in the Czech Republic and Romania.

Read more...