Gaussian decomposition method for full waveform data of LiDAR base on neural network

Area of Science:

• Geomatics Engineering
• Remote Sensing
• Computer Vision

Background:

• Airborne LiDAR full waveform data contains crucial ranging information for point cloud generation.
• Accurate extraction of this ranging data is essential for high-fidelity 3D mapping and modeling.
• Existing methods face challenges in noise reduction and precise parameter estimation from complex waveforms.

Purpose of the Study:

• To introduce an advanced method for Gaussian decomposition of full waveform LiDAR data.
• To improve the accuracy and success rate of extracting ranging information from airborne LiDAR signals.
• To leverage deep learning and refined algorithms for enhanced data processing.

Main Methods:

• Utilized an improved densely connected convolutional neural network (FWDN) for full waveform data preprocessing and noise reduction.
• Employed an enhanced Expectation Maximization (EM) algorithm for Gaussian parameter extraction (amplitude, expectation, FWHM).
• Combined deep learning with statistical methods for robust signal decomposition and ranging information retrieval.

Main Results:

• Achieved a decomposition success rate exceeding 98% on both simulated and measured data.
• Obtained an average range accuracy of less than 1.5 cm, outperforming other existing methods.
• Demonstrated effective noise reduction and precise Gaussian parameter estimation.

Conclusions:

• The proposed method significantly enhances the accuracy of ranging information extraction from airborne LiDAR full waveform data.
• The high success rate and accuracy indicate strong potential for practical applications in mapping and 3D modeling.
• This approach offers a robust solution for processing complex LiDAR waveform data.