Marcus Melander and Alex Thelander will present their master's thesis.

Title: After the Storm: Foundation Models and Classical Baselines for SAR-based Windthrow Detection

Authors: Marcus Melander and Alex Thelander

Supervisor: Anders Heyden

Examinor: Magnus Oskarsson

Abstract:

Wind damage is one of the most economically and ecologically significant natural disturbances in European forests, yet the cloud cover that follows major storms routinely blocks optical satellite acquisitions for weeks, precisely when rapid damage mapping is most valuable. Synthetic aperture radar (SAR) circumvents this limitation through all-weather imaging, but pixel-level windthrow detection from Sentinel-1 remains challenging due to speckle noise, heterogeneous forest structure, and inconsistent labelling regimes across regions. This thesis investigates whether remote sensing foundation models (RSFMs) offer a meaningful advantage over established methods for SAR-based windthrow detection, and whether they constitute a viable approach for operational earth observation. Two multi-modal RSFMs, Galileo and SkySense++, are evaluated against three classical baselines: a backscatter threshold detector, a support vector machine, and a CNN-UPerNet, on four geographically distinct datasets covering Sweden, Russia, Ireland, and Germany. All models share an identical preprocessing pipeline based on five pre-event and five post-event Sentinel-1 acquisitions, and are compared under bi-temporal composite and multi-temporal time-series input configurations, with complementary studies on label efficiency, temporal depth, cross-dataset generalisation, augmentation strategy, and standlevel aggregation. The results show that performance is strongly governed by annotation regime and event geometry rather than by model capacity alone. On the cleanest dataset, SkySense++ reaches 90.1% pixel-level damaged-class F1, substantially exceeding both Galileo and the CNN. On datasets with finer-grained or less consistent annotation, foundation-model gains are modest or negated, and pixel-level performance is bounded primarily by label fidelity. Temporal stacking improves performance on the noisier datasets and plateaus around T = 8 acquisitions, while object-level stand aggregation lifts every learned model well above its pixel-level score and substantially above the threshold baseline, reaching 93.1% object-level F1 on the cleanest dataset. These results suggest that SAR-based windthrow mapping is most credibly deployed as a stand-level prioritisation tool, with foundation models delivering decisive gains where annotation quality permits.