Windthrow mapping in boreal forests using a spatio-temporal deep learning approach and Sentinel-2 imagery

Reliable mapping of windthrow, a common abiotic disturbance in Canadian forests, is crucial for effective forest management and conservation, and remote sensing is becoming a useful tool for this purpose. Due to the spectral similarity of disturbances like windthrow, logging, and insect outbreaks, pixel-based approaches show limitations for accurate windthrow detection at large scales, underscoring the need for a method to distinguish windthrow events across the Canadian boreal forest. In this study, we developed and evaluated a new approach to detect windthrow-affected forest areas, primarily in the boreal forests of Eastern Canada, covering over 90 million hectares from 2019 to 2024. The method integrates deep learning (DL) models with 10 m Sentinel-2 annual composites and incorporates the Continuous Change Detection and Classification (CCDC) algorithm to improve timestamp estimation. We trained and evaluated the performance of machine and DL approaches, namely, LinkNet, CResU-Net, ResU-Net, DeepLabv3+, and Random Forest. Among them, CResU-Net, an enhanced variant of ResU-Net with a convolutional block attention module, demonstrated the best performance, achieving an overall accuracy of 98.79%, a producer’s accuracy of 75.31%, and an Intersection over Union (IoU) of 58.7% using reference data from a major windthrow event in Canada, the May 2022 Canadian derecho. Including CCDC improved our yearly timestamp to monthly timestamp for 63% of the windthrow affected area with a delay of 1–2 months relative to the event. A historical windthrow map from 2019 to 2024 was generated, highlighting the potential of combining DL and Sentinel-2 imagery for accurate and scalable windthrow detection and characterization.