Windthrow detection with moderate to high resolution optical imageries across the European forests

Windthrow is one of the major disturbances for forests and can affect large areas, causing extensive economic and ecological impacts across the different types of forests. Remote sensing is an effective tool with potential to cost-efficiently map large wind-affected regions, with a specific revisit time and spatial resolution, depending on the sensors used. Windthrow detection relies mostly on spatiotemporal changes of forest reflectance. However, one of the main drawbacks in using optical images is the cloud cover, and thus the availability of cloud-free images, especially during the cool season, when most of the windthrow events occur. For this reason, relying solely on the twin satellite Sentinel-2 may not be enough as data with 5-day revisit time are only available since 2016. On the other hand, different approaches have been developed to deploy radar images for scene changes detection, nonetheless, for forests, both L- and C-bands shall be integrated to capture the changes in the different layers of the forest, rendering the process very complicated. In this study, we developed two different approaches for windthrow detection based on the difference between the surface reflectance composite of the image’s (Sentinel-2 and Landsat-8/9) bands, before and after a windthrow event. First, a global machine learning model was developed using multiple windthrow events across Europe in order to classify windthrow events at continental scale. Then, a local machine learning model was developed using samples of damaged areas and non damaged areas of the same forest type in order to classify windthrow events within the same satellite image. Overall, our preliminary results showed that the global model presented relatively lower accuracy and F1 score. This finding is most probably due to the different types of forests, which present different spectral signatures and hamper the correct classification of the affected areas. Conversely, the local model presented higher accuracy and F1 score due to the homogeneity in the selected forest type. Our preliminary results thus indicate that windthrow detection at large scales is still challenging and local models may be a reliable alternative for assessing the wind-affected forests.