Significance The ability of airborne light detection and ranging (LiDAR) to obtain high-precision and high-density three-dimensional (3D) point cloud for a large area has been widely used in many fields such as surveying and mapping, forestry, and the electric power industry. In these applications, detecting specific targets such as buildings, trees, and power lines from the LiDAR point cloud is essential, which is usually considered a classification procedure. However, accurately identifying various typical surface objects from discretely and irregularly distributed LiDAR point clouds is challenging.

Recently, many deep learning-based methods have achieved good performance in airborne LiDAR point cloud classification. Some scholars have employed convolutional neural networks for point cloud classification by transforming the unevenly distributed point cloud into regular images or voxels. In addition, researchers have applied some deep learning-based methods such as PointNet and PointCNN to the original 3D point cloud. In these methods, a large number of training samples are often required; moreover, they are manually labeled, which is time-consuming and laborious, thus inhibiting their wide use in complex scenarios under different conditions. To solve this problem, a small-sample airborne LiDAR point cloud classification method based on integrating transfer learning (TL) and a fully convolutional network (FCN) is proposed in this study.

Progress The proposed classification method is summarized in Fig. 1. First, a colorful airborne LiDAR point cloud is derived using multispectral images; then, the normalized elevation, intensity values, and normalized difference vegetation index of each 3D point cloud are extracted to derive a three-channel point cloud feature map (Fig. 2). Next, multiscale and multiprojection feature maps are constructed by setting different grid sizes(0.1, 0.3, and 0.5 m)(Fig. 3) and projection directions(X, Y, and Z directions)(Fig. 4), which are fed into a pretrained DenseNet201 model to extract deep features of LiDAR point clouds (Fig. 5). Finally, the global features are extracted by applying a max-pooling operation to the extracted deep features for extending the feature description; the max-pooled features are used as the input for an FCN to achieve an initial classification result (Fig. 6), and then the classification result is improved using a graph-cut optimization algorithm.

Results and Discussions Benchmark datasets provided by the International Society of Photogrammetry and Remote Sensing (ISPRS) are used to verify the proposed method. 1) We used four combinations of features to analyze the effects of multiscale and multiprojection features on classification accuracy: a single feature (S1/P1), multiscale features (MS), multiprojection features (MP), and multiscale and multiprojection features (MSMP) (Fig. 8). Experiment results show that the MSMP features had higher per-class F1 scores and overall accuracy than other feature combinations, which could be explained as the MSMP features being able to provide a more effective expression of airborne LiDAR point clouds. 2) To select more suitable pretrained models for classification, we experimentally validated pretrained VGG16, VGG19, ResNet50, and DenesNet201 models (Fig. 9). The contrast result show that the pretrained DenseNet201 model performed better than other pretrained models owing to its ability to extract deep features with better generalization. 3) To validate the effectiveness of graph-cut optimization, we performed graph-cut optimization by setting different numbers of adjacent points K and then chose the best optimization result with the number of adjacent points K=4 for a comparison with the result without graph-cut optimization (Fig. 11). The contrast result show that the graph-cut optimization algorithm could effectively correct misclassification points and improve classification accuracy. 4) The proposed method is compared with methods reported on the ISPRS website and TL-based methods, showing better results in overall accuracy and per-class F1 scores, except for powerline and impervious surfaces, than the methods reported on the ISPRS website. It also achieved better overall accuracy, average F1 score, and per-class F1 scores, except for low vegetation and tree, than the TL-based methods.

In general, the aforementioned experiment results show that the overall accuracy of the proposed method is 89.9% when only per-class 2000 points are used as the training sample (approximately 1.4% of the dataset), which demonstrates that the proposed method can acquire high-precision point cloud classification with fewer training samples and in less time.

Conclusions and Prospect In this study, an airborne LiDAR point cloud classification method based on integrating TL and FCN with small samples is proposed. In the proposed method, multiscale and multiprojection feature maps are first introduced to represent the spatial characteristics of 3D points accurately; then, TL is used to extract deep features from feature maps; finally, the deep features are fed into an FCN for the initial classification of LiDAR point clouds, followed by postprocessing with graph-cut optimization to achieve more accurate classification results. Because many steps are involved in the entire classification process-such as processes of saving and loading data-which prolong the classification time, the entire process will be integrated to accelerate training speed and improve data processing efficiency in the future.