Hierarchical Neural Network Based Data Processing System for Ground- Penetrating Radar. An End of Year Report for CH/1049/6: Application of Neural Networks Coupled With Genetic Algorithms to Optimize Soil Cleanup Operations in Cold Climates

摘要

Ground-Penetrating Radar (GPR) is a powerful modern tool to examine the structure and properties of the media below the ground surface within a depth of 30 meters. This study is very important for the environmental problems related to the transport of contaminants in ground waters. Successful GPR profiling of the subsurface media yielding the correct information about the distribution of permafrost, water table, and bedrock depths is the key factor in ground water flow modeling. This work attempts to develop a hierarchical processing system capable of handling GPR data characterized by high degree of uncertainty, natural physical ambiguity, and, sometimes, missing or incorrect entries. The hierarchical nature of the algorithm allows to split the task of media profiling into several consecutive stages, each of the following has less degree of uncertainty than the previous one. Neural Networks modules are designed to accomplish the two main processing goals of recognizing the 'subsurface pattern' followed by the identification of the depths of the subsurface layers like permafrost, groundwater table, and bedrock. Pre- processing procedure has the objective to transform raw GPR data into a small feature vector containing the most representative and discriminative features of the signal. The feature vector coupled with other relevant GPR information forms the input for the Neural Network processing units.