Improving Forest Damage Detection and Risk Assessment from Winter Storms Using High-Resolution Satellite Data and Environmental Drivers

Abstract. Accurate assessment of forest losses and evaluation of future damage risks are crucial for effective forest management and conservation, particularly as global warming intensifies natural disturbance agents. This study introduces a novel approach combining convolutional neural networks (CNNs) with Random Forest (RF) machine learning classifiers to enhance the precision of forest disturbance detection and risk evaluation. We tested this approach on a large-scale dataset (1490 km²) with diverse forest types and environmental conditions of cool temperate-south boreal forests on Kunashir Island (Northwest Pacific). Using the U-Net deep learning architecture, we precisely identified windthrow patches from (VHR) Pléiades-1 optical satellite imagery. Resulted windthrow map was integrated with an RF classifier that utilized environmental predictors, including elevation, slope aspect, slope inclination, slope curvature, forest canopy closure, landform type, and forest vegetation type, to assess forest damage risk. Our analysis revealed approximately 21.73 km² of the forested area as significantly disturbed, predominantly within dark coniferous forests. Elevation emerged as the most critical predictor of disturbance risk, with complex interactions observed among predictors such as canopy closure and slope steepness. This integrated approach allowed for highly accurate forest loss detection and provided valuable insights into the risk of future damage events. By combining advanced deep learning techniques with RF and detailed environmental predictors, this approach offers a robust framework for evaluating forest disturbance risk. This method pushes forward the frontiers in the precision of forest loss detection but also aids in developing effective strategies for managing and mitigating risks from future disturbance events.