Full-waveform LiDAR echo decomposition based on dense and residual neural networks

Area of Science:

• Geospatial technology
• Signal processing
• Machine learning

Background:

• Conventional full-waveform (FW) LiDAR echo decomposition relies on complex pre-processing algorithms.
• Existing methods face limitations in speed and accuracy for signal analysis.

Purpose of the Study:

• To develop and evaluate highly efficient and accurate decomposition methods for FW LiDAR signals.
• To introduce deep learning approaches, specifically FDCN and FDRN, for improved echo decomposition.

Main Methods:

• Development of a lightweight FW dense connection network (FDCN) for high signal-to-noise ratio (SNR) conditions (>24 dB).
• Implementation of a deeper FW deep residual network (FDRN) with residual blocks for low SNR conditions (e.g., 12 dB).
• Comparative analysis of FDCN and FDRN against conventional decomposition techniques.

Main Results:

• FDCN and FDRN achieve mean errors under 0.2 ns for echo peak location estimation.
• Amplitude error is consistently below 5 mV within a 0-100 mV dynamic range.
• Both deep learning methods demonstrate significantly lower errors compared to conventional approaches.

Conclusions:

• FDCN and FDRN offer superior performance in terms of speed and accuracy for FW LiDAR signal decomposition.
• These deep learning models provide a robust solution for analyzing LiDAR data across various SNR levels.
• The proposed methods represent a significant advancement in LiDAR signal processing and data interpretation.