The global crises – climate change and biodiversity loss – have created a need for precise and wide-scale information of forests. Airborne laser scanning (ALS) provides a means for collecting such information, as it enables mapping large areas efficiently with a resolution sufficient for object-level information extraction. Deadwood is an important component of the forest environment, as it stores carbon and provides a habitat for a wide variety of species. Mapping deadwood provides information about the valuable areas regarding biodiversity, which can be used in, e.g., conservation and restoration planning. The aim of this thesis was to develop automated methodology for detecting individual fallen and standing dead trees from ALS data.
Studies I and II presented a line detection based method for detecting fallen trees and evaluated its performance on a moderate-density ALS dataset (point density approx. 15 points/m2) and a high point density unmanned aerial vehicle borne laser scanning (ULS) dataset (point density approx. 285 points/m2). In addition, the studies inspected the dataset, methodology, and forest structure related factors affecting the performance of the method. The studies found that the length and diameter of fallen trees significantly impact their detection probability, and that the majority of large fallen trees can be identified from ALS data automatically. Furthermore, study I found that the amount and type of undergrowth and ground vegetation, as well as the size of surrounding living trees determine how accurately fallen trees can be mapped from ALS data. Moreover, study II found that increasing the point density of the laser scanning dataset does not automatically improve the performance of fallen tree detection, unless the methodology is adjusted to consider the increase in noise and detail in the point cloud.
Study III inspected the feasibility of high-density discrete return ULS data for mapping individual standing dead trees. The individual tree detection method developed in the study was based on a three-step process consisting of individual tree segmentation, feature extraction, and machine learning based classification. The study found that, while some of the large standing dead trees could be identified from the ULS dataset, basing detection on discrete return data and the geometrical properties of trees did not suffice for acquiring applicable deadwood information. Thus, spectral information acquired with multispectral laser scanners or aerial imaging, or full-waveform laser scanning is necessary for detecting individual standing dead trees with a sufficient accuracy.
The findings of this thesis contribute to the existing deadwood detection methodology and improve the understanding of factors to take into account when utilizing ALS for detecting dead trees at a single-tree-level. Although remotely sensed deadwood mapping is still far from a resolved topic, these contributions are a step towards operationalizing remotely sensed biodiversity monitoring.