Hybrid forest disturbance classification using Sentinel-1 and inventory data: a case-study for Southeastern USA

Abstract

Forest ecosystems are increasingly threatened by disturbances such as fires, droughts, storms, and insect and pathogen outbreaks. Accurate and timely disturbance mapping is essential for understanding their dynamics and informing mitigation strategies to combat widespread forest decline. Traditional inventories, such as the U.S. Forest Service’s Insect and Disease Survey (IDS), provide detailed information on biotic and abiotic disturbances; however, they have varying coverage and inherent uncertainties in the location, extent, and timing of disturbances due to data-collection constraints. Other approaches, such as satellite remote sensing, can, in principle, overcome some of these challenges by providing large-scale coverage and continuous spatio-temporal observations. However, robust disturbance classification algorithms need to be developed, which in turn require good-quality labels. We present a novel approach for refining disturbance classification labels by combining IDS with Sentinel-1 radar backscatter change detection to produce a new reference dataset, Sentinel-1 Disturbance Mapping (S1DM). The disturbed patches identified by Sentinel-1 are typically located within 200–330 m of IDS locations and generally agree on disturbance timing. Sentinel-1 tends to detect bark beetle disturbances up to 2 years earlier than IDS, and some defoliator events are also detected 1 year prior to IDS. When statistically examined against manual labels from high-resolution images from PlanetScope, we found that S1DM performed better than IDS for wind and bark beetle disturbances, but not for defoliators. For bark beetle disturbances, S1DM improves the median Intersection over Union from 0.007 to 0.076, an absolute gain of 6.9 % over IDS. By integrating spatial and temporal information on disturbance occurrence from Sentinel-1 change detection with information on the corresponding disturbance agent from IDS, S1DM provides a high-quality forest disturbance reference dataset for developing remote sensing forest classification models. Our approach highlights the benefits of combining satellite-based remote sensing with traditional aerial survey data, reducing aerial survey costs while providing a scalable method adaptable to various regions.