Point cloud semantic segmentation, a challenging task in 3D data processing, is popular in many realistic applications. Currently, deep learning methods are gradually being applied to point cloud semantic segmentation. However, as it is difficult to manually label point clouds in 3D scenes, it remains difficult to obtain sufficient training samples for the supervised deep learning network. Although an increasing number of excellent methods have been proposed in recent years, few of these have focused on the problem of semantic segmentation with insufficient samples. To address this problem, this paper proposes a random graph based on graph convolution network, referred to as RG-GCN. The proposed network consists of two key components: (1) a random graph module is proposed to perform data augmentation by changing the topology of the built graphs; and (2) a feature extraction module is proposed to obtain local significant features by aggregating point spatial information and multidimensional features. To validate the performance of the RG-GCN, the indoor dataset S3DIS and outdoor dataset Toronto3D are used to validate the proposed network via a series of experiments. The results show that the proposed network achieves excellent performance for point cloud semantic segmentation of the two different datasets.

Canopy cover is an important indicator and commonly used in forest management applications. Unmanned-Aerial-Vehicle (UAV)—Borne Laser Scanning (ULS) has drawn increasing attention as a new alternative source for forest field inventory due to its spatial resolution comparable to that of Terrestrial Laser Scanning (TLS). In this study, the performance of plot canopy cover estimations from ULS and TLS is investigated. The experiment was conducted in 16 plots from two Pinus massoniana forests with different stand conditions in Guangxi, China. Both the Canopy Height Model (CHM)-based and Individual Tree Delineation (ITD)-based methods were used to estimate the canopy cover. The influence of CHM pixel sizes on the estimations was also analyzed. Our results demonstrated that the accuracies of ULS ($R^2$: 0.992–0.996, $RMSE$: 0.591–0.820%) were better than those of TLS ($R^2$: 0.541–0.846, $RMSE$: 3.642–6.297%) when compared against the reference. The average difference between the ULS and TLS estimations was 6.91%, and the disagreement increased as the forest complexity increased. The reasonable CHM pixel sizes for the canopy cover estimations were 0.07–1.2 m for ULS and 0.07–1.5 m for TLS. This study can provide useful information for the selection of data sources and estimation methods in plot canopy cover mapping.

The current status as well as the potential absorption capability of the forest carbon sinks in terrestrial ecosystems are in urgent need of further studies. The ground observation-based methods are labor-intensive, and the resulting statistics from samples are difficult to evaluate, while inversion methods based on remote sensing data lack theoretical explanation and universality. This paper proposes a pixel-level, multi-scale, high-precision Explicit Forest carbon stock Model (EFM) that is universal and adaptive. First, four key variables were used in the construction of the EFM: remote sensing image resolution, forest canopy density, terrain slope, and forest height; Second, simulated forest scene were generated based on the growth characteristics of individual trees, and EFM parameters were solved and analyzed by these simulated pixels; Third, the EFM was tested at various scales and forest saturation levels to verify its accuracy, robustness, and applicability, the simulation and real-life experiments show that the correlation coefficient is greater than 0.98 and the relative error is about 20%. The EFM solves the problem that the existing methods are lack in theoretical interpretation and universal applicability, thus can be used to map forest carbon stocks at high resolution and large scale and even monitor forest carbon dynamics at global scale.