Automated storm damage severity mapping from satellite imagery using image segmentation and machine learning

Abstract:

There is evidence that climate change is increasing the frequency and severity of storms. In forested landscapes, mapping storm damage severity is important for multiple applications, including rescue and response resource allocation, prioritizing salvage areas, and assessing impacts on habitat and ecosystem function. Yet the creation of damage severity maps can be challenging, especially when correlating “severity” as depicted in imagery to measurements of “severity” that a ground observer (i.e., the user or responder) might make. Generally, two issues arise: 1) image measurements may not match the needs of the user in terms of spatial, temporal or spectral resolution, and, 2) classifying image measurements based upon spectral rather than biological terms may not translate well into usable information on the ground. Satellites such as Sentinel-2 and Landsat may have long revisit times when cloud cover is problematic, and recovering vegetation may obscure the effects of a disturbance. Lower latency solutions such as airplane or drone data may not be practical given weather, spatial extent or cost. In addition, custom datasets from aircraft can be problematic for responders because users may not have the time or experience to process the image rapidly to respond to a disturbance. Here, we describe a suite of solutions to these challenges, offering multiple approaches for differing storm types and response needs, combining high-temporal frequency Planet Dove imagery, radiometrically and spectrally robust Sentinel-2 and Landsat imagery, and the XGBoost machine learning algorithm to map storm damage severity. We show how Planet and Sentinel-2 imagery can be used to map windstorm damage into categories usable by responders for both near-real-time applications and on a longer time scale for recovery and salvage harvest efforts, and fire risk assessment. To facilitate this critical capability for resource-limited response agencies, these methods are automated and designed to generate easy-to-use ESRI shapefiles that a user can employ to guide in-field triage assessment. Written entirely in open-source Python, this work aims to greatly improve the resources available to those responding to storm events, providing them with tools to focus their efforts where needed most.