Lidar detection of underwater objects using a neuro-SVM-based architecture

Area of Science:

• Remote Sensing
• Machine Learning
• Signal Processing

Background:

• Light detection and ranging (Lidar) data provides backscatter intensities from airborne objects.
• Existing Lidar data classification methods struggle with noise and irrelevant air backscatter information.
• Accurate classification is crucial for detecting underwater objects using Lidar.

Purpose of the Study:

• To develop and evaluate a novel neural network architecture for Lidar data classification.
• To improve the accuracy of underwater object detection from Lidar data.
• To compare the proposed architecture against traditional classifiers like Bayesian and quadratic classifiers.

Main Methods:

• Pre-filtering Lidar data to remove high-frequency noise and air backscatter.
• Feature extraction using linear predictive coding (LPC) and polynomial approximation.
• A parallel neural network architecture (MLP and HRBF) feeding into an SVM inference engine.

Main Results:

• The proposed parallel neural network architecture achieved a high prediction accuracy of 98.9%.
• The architecture effectively handles noise and irrelevant air backscatter in Lidar data.
• Bayesian, quadratic, and single-layered ANN classifiers demonstrated lower prediction accuracy.

Conclusions:

• The developed parallel neural network architecture with an SVM inference engine is highly suitable for Lidar data classification.
• This approach significantly enhances the accuracy of underwater object detection.
• The proposed method outperforms traditional classification techniques for this specific task.