Errors related to the automatized satellite-based change detection of boreal forests in Finland

Highlights • Forest changes were automatically modelled from multitemporal Sentinel-2 images. • Errors were evaluated based on visually interpreted VHR images. • Extraction of clear-cuts was accurate whereas thinnings had more variation. • Image quality and translucent clouds had most significant effect on errors. • Results were regarded applicable for operational change monitoring.

The majority of the boreal forests in Finland are regularly thinned or clear-cut, and these actions are regulated by the Forest Act. To generate a near-real time tool for monitoring management actions, an automatic change detection modelling chain was developed using Sentinel-2 satellite images. In this paper, we focus mainly on the error evaluation of this automatized workflow to understand and mitigate incorrect change detections. Validation material related to clear-cut, thinned and unchanged areas was collected by visual evaluation of VHR images, which provided a feasible and relatively accurate way of evaluating forest characteristics without a need for prohibitively expensive fieldwork. This validation data was then compared to model predictions classified in similar change categories. The results indicate that clear-cuts can be distinguished very reliably, but thinned stands exhibit more variation. For thinned stands, coverage of broadleaved trees and detections from certain single dates were found to correlate with the success of the modelling results. In our understanding, this relates mainly to image quality regarding haziness and translucent clouds. However, if the growing season is short and cloudiness frequent, there is a clear trade-off between the availability of good-quality images and their preferred annual span. Gaining optimal results therefore depends both on the targeted change types, and the requirements of the mapping frequency.