Deep learning-based tree classification using mobile LiDAR data

Abstract

Our work addresses the problem of extracting and classifying tree species from mobile LiDAR data. The work includes tree preprocessing and tree classification. In tree preprocessing, voxel-based upward-growing filtering is proposed to remove ground points from the mobile LiDAR data, followed by a tree segmentation that extracts individual trees via Euclidean distance clustering and voxel-based normalized cut segmentation. In tree classification, first, a waveform representation is developed to model geometric structures of trees. Then, deep learning techniques are used to generate high-level feature abstractions of the trees' waveform representations. Quantitative analysis shows that our algorithm achieves an overall accuracy of 86.1% and a kappa coefficient of 0.8 in classifying urban tree species using mobile LiDAR data. Comparative experiments demonstrate that the uses of waveform representation and deep Boltzmann machines contribute to the improvement of classification accuracies of tree species.

The Use of Three-Dimensional Convolutional Neural Networks to Interpret LiDAR for Forest Inventory

“…A few studies have begun using deep learning for measuring and analyzing forest attributes. For example, Guan et al [34] used a segmentation technique to isolate tree crowns, and then used a neural network to classify species based on point distribution. In another study, Ghamisi et al [35] applied a 2D CNN to estimate forest attributes from rasterized LiDAR and hyperspectral data.…”

Pole-Like Road Object Detection in Mobile LiDAR Data via Supervoxel and Bag-of-Contextual-Visual-Words Representation

“…To generate bag-of-contextual-visual-words representations from the filtered off-ground points, individual objects are first obtained through the following steps: 1) a Euclidean distance clustering approach, with a clustering distance of d c , is applied, followed by a voxel-based normalized cut (Ncut) segmentation method [7]. In [7], the Ncut method effectively segments connected, but not seriously overlapped, clusters into separated semantic objects.…”

“…To reduce the number of points to be processed, a voxelbased upward growing filtering method [7] is first performed to remove ground points from the test data set. Fig.…”

“…becoming a leading source for highway mapping [3], urban road distress assessment [4], [5], and road feature inventory [6], [7]. Pole-like road objects, including light poles and traffic signposts, located along roads/streets, are typical kinds of road infrastructure.…”

“…In [11], a pairwise 3-D shape context descriptor, which considers both local and global similarity measures, was proposed to detect light poles. In [7], a 3-D object matching framework was proposed for detecting light poles, with attachments, of varied shapes and sizes.…”