An unsupervised classification-based time series change detection approach for mapping forest disturbance

Abstract

Unsupervised Classification to Change (UC-Change) is a new remote sensing approach for mapping areas affected by logging and wildfires. It addresses the main limitations of existing image time-series change detection techniques, such as limited multi-sensor capabilities, use of purely spectral-based forest recovery metrics, and poor detection of salvage harvesting. UC Change detects disturbances and tracks forest recovery by analyzing changes in the spatial distribution of spectral classes over time. The algorithm detected approximately 85% and 70% of reference cutblock and fire scar pixels at a ±2-year temporal agreement, respectively, consistently outperforming existing algorithms across different biogeoclimatic zones of British Columbia, Canada. The results indicate an upper estimate of 7.5 million ha of forest cleared between 1984 and 2014, which is above estimates based on existing maps and databases (6.3 – 6.7 million ha). Also presented is a new framework for using open-access data for validation of change detection results.