ALS Point Cloud Classification by Integrating an Improved Fully Convolutional Network into Transfer Learning with Multi-Scale and Multi-View Deep Features

Airborne laser scanning (ALS) point cloud has been widely used in various fields, for it can acquire three-dimensional data with a high accuracy on a large scale. However, due to the fact that ALS data are discretely, irregularly distributed and contain noise, it is still a challenge to accurately identify various typical surface objects from 3D point cloud. In recent years, many researchers proved better results in classifying 3D point cloud by using different deep learning methods. However, most of these methods require a large number of training samples and cannot be widely used in complex scenarios. In this paper, we propose an ALS point cloud classification method to integrate an improved fully convolutional network into transfer learning with multi-scale and multi-view deep features. First, the shallow features of the airborne laser scanning point cloud such as height, intensity and change of curvature are extracted to generate feature maps by multi-scale voxel and multi-view projection. Second, these feature maps are fed into the pre-trained DenseNet201 model to derive deep features, which are used as input for a fully convolutional neural network with convolutional and pooling layers. By using this network, the local and global features are integrated to classify the ALS point cloud. Finally, a graph-cuts algorithm considering context information is used to refine the classification results. We tested our method on the semantic 3D labeling dataset of the International Society for Photogrammetry and Remote Sensing (ISPRS). Experimental results show that overall accuracy and the average F1 score obtained by the proposed method is 89.84% and 83.62%, respectively, when only 16,000 points of the original data are used for training.