Mapping old-growth forests using airborne lidar data and satellite images: how do plot size and rarity affect accuracy?

Abstract

Old-growth forests have become rare and fragmented in the boreal biome. Their precise locations are not currently known with sufficient accuracy to support forest conservation and forest management. We studied the mapping of old-growth forests using airborne lidar data and satellite images in the Finnish coniferous forests. We investigated how plot size and the rarity of old-growth forests affect the accuracy of old-growth forest detection. We employed a Gaussian process classifier to distinguish old-growth forests from managed forests. Our field data consisted of 176 old-growth and 1082 managed forest plots. The results showed that an increase in plot size from 20 m × 20 m to 60 m × 60 m improved the performance of the classifier, because the larger plots more likely contain spatial patterns of trees and crown features indicative of forest naturalness. The largest F1-score (0.74) was achieved by data augmentation that generates additional training plots located inside forest boundaries. We also showed that the detection accuracy of old-growth forests decreases as they become rarer in the population. This rarity effect is crucial to understand, because the occurrence of old-growth forests can vary regionally due to different land use pressures. The mapping procedure proposed here can assist in the planning of field-based inventories of old-growth forests.