Detection of snow disturbances in boreal forests using unitemporal airborne lidar data and aerial images

Snow is among the most significant natural disturbance agents in Finland. In silviculture, maps of snow disturbance are needed to recognize severely disturbed forests where the risk of subsequential disturbances, such as insect outbreaks, is high. We investigated the potential of unitemporal airborne lidar (light detection and ranging) data and aerial images to detect snow disturbance at the tree level. We used 81 healthy and 128 snow-disturbed field plots established in a 63 800 ha study area in Eastern Finland. A subset of trees (n = 675) was accurately positioned in the field plots. We carried out individual tree detection (ITD) using airborne lidar data (5 p/m2), and a random forest classifier was used to classify healthy and broken trees. Tree features were extracted from a terrain elevation model, lidar data, and aerial imagery. We compared canopy height model–based (ITDCHM) and point cloud–based (ITDPC) ITD approaches. We explored random forest variable importance scores and evaluated the classification performance by an F1-score and its components (precision and recall). Performance was also evaluated at the plot level to investigate errors associated with the predicted number of broken trees. We achieved F1-scores of 0.66 and 0.85 for the tree- and plot-level classifications, respectively. The variable importance scores showed that elevation above sea level was the most important predictor variable followed by ITD-based features characterizing the neighborhood of trees. The ITDCHM slightly outperformed the ITDPC at the tree level, while they both underestimated the number of broken trees at the plot level. The proposed approach can be carried out alongside lidar-assisted operational forest management inventories provided that a set of positioned broken and healthy trees are available for model training. Since airborne lidar data often have a temporal resolution of several years for the same areas, future research should consider the utilization of other remotely sensed data sources to improve the temporal resolution.