Student name: Pirkko Väkimies

RUN-EU institution: Häme University of Applied Sciences, Finland

Abstract

Deadwood plays a significant role in maintaining biodiversity by providing food and habitats for numerous species, as well as influencing the carbon cycle. A review of research articles related to the identification of fallen deadwood from ALS LiDAR data using machine learning, particularly convolutional neural networks, was found to be a relatively unexplored topic.

The aim of this research was to explore machine learning methods and their combinations to detect fallen deadwood from point cloud data. As a starting point, scientific research articles related to deadwood identification were reviewed. Synthesis based on these articles, led to a use of LiDAR data (5 p/m²) and Hough-transformation as well as convolutional neural network. Field measurements were conducted to verify the methods. LiDAR data was filtered by removing points from higher vegetation. The filtered data was used to create four different point cloud datasets (pixel images) by using different KNN-methods.

Hough transformations were applied to these datasets using eight different parameterisations to predict potential fallen deadwood. The Hough-based predictions were run using a convolutional neural network, in hope to refine them, by utilising TensorFlow and the Adam optimization algorithm. The obtained results were compared to field measurements.

A new metric called “deadwood occurrence density” [m/m²] was calculated to describe the potential of deadwood in an area. In other words, determining the number of precisely identified fallen deadwood was not extremely important for the overall result – instead, the focus was on whether it was likely, based on the metric, that deadwood would be present in the area. The “probability of deadwood occurrence” was derived from the gamma distribution of the predicted occurrence density. Likelihood was categorised into three classes: unlikely (1), slightly likely (2), and likely (3). The resulting likelihood classes were visualised on a one-square-kilometer (1 km²) map layout, with colour-coded 50 x 50 m (2500 m²) grid cells to help with field verification of potential deadwood locations.

The Hough transformation predictions were closest to the field-measurements. As a result of the derived metric (probability of deadwood occurrence), the study was able to predict areas with deadwood, as well as null block where no fallen deadwood was observed. However, the method studied is not foolproof – the method itself does not prove anything whether the observation is deadwood or freshly fallen trunks. The method can be extended with new deadwood-predicting indicators, such as identifiers of large deciduous tree crowns.

I strongly believe that the investigated method should be developed further. Especially as we recognise the urgency in mapping biodiversity-rich areas, such as old forests and areas containing significant amounts of deadwood.

Keywords: Hough-transformation, Knn-method, convolutional neural network, fallen dead wood, LiDAR data, Python, QGIS.