Improved Supervised Learning-Based Approach for Leaf and Wood Classification From LiDAR Point Clouds of Forests

Abstract

Accurately classifying 3-D point clouds into woody and leafy components has been an interest for applications in forestry and ecology including the better understanding of radiation transfer between canopy and atmosphere. The past decade has seen an increase in the methods attempting to classify leaves and wood in point clouds based on radiometric or geometric features. However, classification purely based on radiometric features is sensor-specific, and the method by which the local neighborhood of a point is defined affects the accuracy of classification based on geometric features. Here, we present a leaf-wood classification method combining geometrical features defined by radially bounded nearest neighbors at multiple spatial scales in a machine learning model. We compared the performance of three different machine learning models generated by the random forest (RF), XGBoost, and lightGBM algorithms. Using multiple spatial scales eliminates the need for an optimal neighborhood size selection and defining the local neighborhood by radially bounded nearest neighbors makes the method broadly applicable for point clouds of varying quality. We assessed the model performance at the individual tree-and plot-level on field data from tropical and deciduous forests, as well as on simulated point clouds. The method has an overall average accuracy of 94.2% on our data sets. For other data sets, the presented method outperformed the methods in literature in most cases without the need for additional postprocessing steps that are needed in most of the existing methods. We provide the entire framework as an open-source python package.Index Terms-Leaf versus wood separation, LiDAR, machine learning, python package, tropical forests.

An Improved Convolution Neural Network-Based Model for Classifying Foliage and Woody Components from Terrestrial Laser Scanning Data

“…The point density is another factor that can affect the classification results in similar researches [9,10,15]. However, the results of this study verified that the classification results obtained using the FWCNN model are insensitive to the point density variation of tested TLS datasets.…”

“…where P 0 means the relative observed agreement among N samples in error matrices with r rows, which is equal to the ratio of the sum of the correct identification samples of all samples belonging to the same type; the P c is the hypothetical probability of chance agreement; X i+ and X +i are the row probabilities and the column probabilities, respectively. To assess the contribution of combining the LRI and geometrical information in distinguishing TLS data, we used the LeWoS model [32] and LWCLF model [15], which are only based on geometrical features, to classify the TLS datasets, and evaluated their classification accuracy for foliage and woody components. Additionally, we compared the FWCNN-based results with those obtained using the Random Forest (RF) algorithm [48], Gaussian Mixed Model (GMM) [30], and Support Vector Machine (SVM) algorithm [49] in terms of classification accuracy and running efficiency.…”

“…Two commonly used approaches for separating foliage and woody points from TLS data are the local geometrical feature- [8,9,12] and LRI-based methods [11,13]. The linear (e.g., branches and twigs), surface (e.g., broadleaves), and random (e.g., needle clusters) distribution of local points in TLS data can be used to distinguish foliage and woody points [9,14,15]. However, it is still challenging to classify TLS data in a natural forest based on geometrical features alone due to the complex forest 3-D structure resulting from the mixed tree species and the interlacing branches and foliage within the tree crowns.…”

“…The geometrical and LRI-related features of local points in TLS data could be extracted by various methods-such as the sphere searching-based [24], voxel-based [25,26], patch-based [27], KD tree-based [28], or K nearest-based [29] methods-for classifying woody and foliage points. Among these methods, the sphere searching method is a well-accepted and broadly applied method due to its robustness [15,24] in keeping the local structural feature of discrete point cloud data. However, the selection of the optimal radius of searching sphere should be carefully conducted for better local feature extraction.…”

“…Existing deep learning models, such as PointNet [33] and PointNet++ [24], were based on global or local geometrical features to classify point cloud data. Moreover, some models were tested to extract multi-scale local geometrical features from point cloud data to improve classification accuracy [15,36,37]. However, the above deep learning models did not use the LRI-based features to classify and segment point cloud data.…”