Ground-Truthing Forest Change Detection Algorithms in Working Forests of the US Northeast

Abstract

The need for reliable landscape-scale monitoring of forest disturbance has grown with increased policy and regulatory attention to promoting the climate benefits of forests. Change detection algorithms based on satellite imagery can address this need but are largely untested for the forest types and disturbance regimes of the US Northeast, including management practices common in northern hardwoods and mixed hardwood-conifer forests. This study ground-truthed the “off-the-shelf” outputs of three satellite-based change detection algorithms using detailed harvest records and maps covering 43,000 ha of working forests in northeastern New York. Study Implications: Algorithms performed best in detecting clearcuts, but performed much worse and poorly overall in detecting the partial harvest prescriptions (e.g., shelterwoods, thinnings) that were far more common in our ground-truthing data (and for this region). Among the algorithms tested, Landtrendr was consistently superior at both detecting partial harvests and estimating harvest intensity (volume removals), but there still remained substantial room for improvement. Overall, we suggest that these algorithms need further training and tuning to be reliably used for accurate monitoring of harvest-related activities in working forests of the US Northeast.