Deep Learning for LiDAR Point Cloud Classification in Remote Sensing.

Area of Science:

• Computer Vision
• Remote Sensing
• Geospatial Data Analysis

Background:

• Point clouds are a primary data format from depth sensors, crucial for 3D data analysis.
• Deep learning (DL) methods have shown significant success in computer vision tasks like classification and segmentation of 3D point cloud data.
• Existing research on DL for point clouds in remote sensing lacks a comprehensive roadmap, highlighting limitations and challenges.

Purpose of the Study:

• To provide a roadmap of state-of-the-art deep learning models for point cloud processing in remote sensing.
• To categorize DL models based on their data structure input.
• To benchmark model performance on standard datasets and summarize available 3D datasets for DL.

Main Methods:

• Categorization of deep learning models by input data structure.
• Performance evaluation and benchmarking of models on widely used remote sensing datasets.
• Comparative analysis of different DL architectures for point cloud analysis.

Main Results:

• Convolutional Neural Networks (CNNs) demonstrate the highest performance across various remote sensing applications.
• Light-weighted CNN models, specifically Dynamic Graph CNN (DGCNN) and ConvPoint, offer excellent performance.
• A summary of benchmark 3D datasets suitable for deep learning training and testing is provided.

Conclusions:

• CNNs are highly effective and efficient for remote sensing tasks utilizing point cloud data.
• DGCNN and ConvPoint represent leading models for state-of-the-art performance in this domain.
• The study offers valuable insights for researchers and practitioners in applying DL to remote sensing point clouds.