Airborne light detection and ranging (LiDAR) technology has been widely utilized for collecting three-dimensional (3D) point cloud data on forest scenes, enabling the generation of high-accuracy digital elevation models (DEMs) for the efficient investigation and management of forest resources. Point cloud filtering serves as the crucial initial step in DEM generation, directly influencing the accuracy of the resulting DEM. However, forest filtering presents challenges in dealing with sparse point clouds and selecting appropriate initial ground points. The introduction of full-waveform LiDAR data offers a potential solution to the problem of sparse point clouds. Additionally, advancements in multi-source data integration and machine learning algorithms have created new avenues that can address the issue of initial ground point selection. To tackle these challenges, this paper proposes a novel filtering method for forest scenes utilizing full-waveform LiDAR data and hyperspectral image data. The proposed method consists of two main steps. Firstly, we employ the improved dynamic graph convolutional neural network (IDGCNN) to extract initial ground points. In this step, we utilize three types of low-correlation features: LiDAR features, waveform features, and spectral features. To enhance its accuracy and adaptability, a self-attention module was incorporated into the DGCNN algorithm. Comparative experiments were conducted to evaluate the effectiveness of the algorithm, demonstrating that the IDGCNN algorithm achieves the highest classification accuracy with an overall accuracy (OA) value of 99.38% and a kappa coefficient of 95.95%. The second-best performer was the RandLA-net algorithm, achieving an OA value of 98.73% and a kappa coefficient of 91.68%. The second step involves refining the initial ground points using the cloth simulation filter (CSF) algorithm. By employing the CSF algorithm, non-ground points present in the initial ground points are effectively filtered out. To validate the efficacy of the proposed filtering method, we generated a DEM with a resolution of 0.5 using the ground points extracted in the first step, the refined ground points obtained with the combination of the first and second steps, and the ground points obtained directly using the CSF algorithm. A comparative analysis with 23 reference control points revealed the effectiveness of our proposed method, as evidenced by the median error of 0.41 m, maximum error of 0.75 m, and average error of 0.33 m.

The existing deep learning-based Airborne Laser Scanning (ALS) point cloud semantic segmentation methods require a large amount of labeled data for training, which is not always feasible in practice. Insufficient training data may lead to over-fitting. To address this issue, we propose a novel Multi-branch Feature Extractor (MFE) and a three-stage transfer learning strategy that conducts pre-training on multi-source ALS data and transfers the model to another dataset with few samples, thereby improving the model’s generalization ability and reducing the need for manual annotation. The proposed MFE is based on a novel multi-branch architecture integrating Neighborhood Embedding Block (NEB) and Point Transformer Block (PTB); it aims to extract heterogeneous features (e.g., geometric features, reflectance features, and internal structural features) by leveraging the parameters contained in ALS point clouds. To address model transfer, a three-stage strategy was developed: (1) A pre-training subtask was employed to pre-train the proposed MFE if the source domain consisted of multi-source ALS data, overcoming parameter differences. (2) A domain adaptation subtask was employed to align cross-domain feature distributions between source and target domains. (3) An incremental learning subtask was proposed for continuous learning of novel categories in the target domain, avoiding catastrophic forgetting. Experiments conducted on the source domain consisted of DALES and Dublin datasets and the target domain consists of ISPRS benchmark dataset. The experimental results show that the proposed method achieved the highest OA of 85.5% and an average F1 score of 74.0% using only 10% training samples, which means the proposed framework can reduce manual annotation by 90% while keeping competitive classification accuracy.