Skip to main content
Dryad logo

Detection of standing retention trees in boreal forests with airborne laser scanning point clouds and multispectral imagery

Citation

Hardenbol, Alwin; Korhonen, Lauri; Kukkonen, Mikko; Maltamo, Matti (2022), Detection of standing retention trees in boreal forests with airborne laser scanning point clouds and multispectral imagery, Dryad, Dataset, https://doi.org/10.5061/dryad.fqz612jw3

Abstract

1. In a landscape consisting primarily of intensive forestry interspersed with some protected areas, multifunctional forestry with retention trees can play a crucial role in nature conservation. Accurate mapping of retention trees is important for guiding landscape-level conservation and forest management and improving landscape connectivity. Sizeable dead and living retention trees play a particularly important ecological role but even their large-scale inventory is often intensive through field work and/or inaccurate. We aimed to detect and classify retention trees using the novel nationwide Finnish airborne laser scanning (ALS) data (~ 5 pulses/m2) in conjunction with unrectified color-infrared (CIR) aerial imagery. 2. Applying photogrammetric principles, we added spectral information from the CIR imagery to the ALS-derived point cloud. For a training dataset of 160 retention trees from 19 stands and a geographically separate validation dataset of 79 trees from 8 stands, we segmented trees via individual tree detection (ITD), removed most trees belonging to the regenerating vegetation layer, and classified trees into living conifers, living broadleaves, and dead trees by linear discriminant analysis. 3. The detection rate via ITD differed considerably for dead and living trees, with 41.7% of all dead and 83.8% of all living trees being detected with relatively low commission error rates. Dead trees with smaller diameters and heights were more likely missed, while grouping caused living tree omission. For classification into living conifers, living broadleaves, and dead trees, an overall accuracy of 67.3% was achieved in training and 71.2% in validation data only ALS-derived metrics. When adding spectral metrics, the overall accuracies were 79.6% and 61.0% for training and validation, respectively. 4. Our findings imply that wall-to-wall large-scale high density ALS data can be used to detect retention trees rather accurately – even larger dead trees – and that metrics derived solely from ALS data can accurately classify detected retention trees into living conifers, living broadleaves, and dead trees. Considering the ecological value of retention trees, our results are promising and indicate that ALS data of the studied pulse density are a cost-effective option for large area mapping of retention trees in countries with such data available.

Methods

The field data was collected as followed: Two field inventories were performed to gather data for a training dataset (collected in May 2021) and a validation dataset (collected in October 2021) for which no permits were required. To include a sufficient number of standing dead trees in both datasets, we mostly inventoried stands that had at least one standing dead tree. We define retention trees as trees that stuck out at least three meters from the regenerating vegetation layer. We included stands that underwent final harvesting at any time in the past and the regenerating vegetation layer thus ranged from 0 to 13 m (recent to old final felling).

All standing retention trees above 9 cm in DBH were inventoried. We selected 9 cm DBH as a cut-off point because it is reasonable to assume that thinner dead trees are likely missed frequently by the laser pulses. For suitable retention trees, we recorded tree species, conditions (dead or living), positions, heights, diameters, and only for dead trees, stem and crown conditions. Tree species that were recorded were Norway spruce, Scots pine, silver birch, downy birch, European aspen, rowan, goat willow, and grey alder. Note, however, that we grouped broadleaves and conifers to obtain good samples sizes for the classification. Positions were recorded with a Trimble Geo 7x Global Navigation Satellite Systems handheld and a Trimble Tornado antenna (Trimble, California, US) at an accuracy of < 1 m for horizontal coordinates. With the help of publicly available CIR orthophotos provided by the National Land Survey of Finland, the tree positions were verified where visible and corrected if required. Tree heights were measured with a Haglöf Vertex IV and the accompanying Transponder T3 (Haglöf, Långsele, Sweden). Diameters were measured in two distinct ways. If a tree was intact, DBH was measured at 1.3 m height above ground. If a tree was broken, diameter was estimated at the middle height of the remaining standing tree column. For dead trees only, we recorded if the stem was intact or broken (either naturally or artificially), and if the stem was intact, whether the crown was entirely present, partially present, or absent. In total, we collected field data on 160 retention trees from 19 stands for the training dataset and on 79 retention trees from 8 stands for the validation dataset.

The two remotely sensed datasets [airborne laser scanning (ALS) data with an average pulse density of 5 pulses/m2 and unrectified color-infrared aerial (CIR) imagery] that we used were collected by third parties. Note that we cannot share the raw ALS data and unrectified CIR images on Dryad for legal reasons. We processed these data by applying photogrammetric principles and thus added spectral information (red, green, and near infrared bands) from the unrectified multispectral imagery to the ALS-derived point cloud. This was done through collinearity equations with knowledge of camera locations and orientations at the time of exposure of each image and internal camera parameters. The per channel digital number value for each pixel was averaged from all the images where a 3D point was observed. The result of this process was a point cloud that included both XYZ-coordinates and spectral information for each return. We have obtained permission to share these ALS data with already attached spectral values of all stands included in our analyses. A final processing step that we have already applied to these .laz files are height normalization, setting heights of points that were negative to zero, and removing outliers (high height values).

Usage Notes

R (version 3.5.3 but later versions such as 4.2.1 also worked) and LAStools (research/commercial license) are required. Additionally, a GIS like QGIS is useful for visualisation (starting from QGIS version 3.26 there is also good support for .las/.laz files).

Funding

Academy of Finland Flagship Programme (Forest-Human-Machine Interplay - Building Resilience, Redefining Value Networks and Enabling Meaningful Experiences (UNITE), Award: 337127