5:15 PM - 6:45 PM
[HTT18-P01] 3D interpretation support tool for quality control of automatic interpretation results of underground radar data using deep learning
Keywords:ground-penetrating radar, machine learning, interpretation, automation
In underground radar analysis, skilled engineers visually decipher the shape and spatial distribution of reflected waves due to differences in permittivity of underground structures from cross-sectional images obtained from moving vehicles. In recent years, advances in data collection system technology have led to the acquisition of large amounts of data, making it essential to automate and save labor in interpretation. Furthermore, many model architectures have been proposed in machine learning technology, and the ability to recognize objects has improved dramatically. Systems that apply these new technologies to automatically identify ground-penetrating radar data have been developed and implemented; some are in practical use. As a result, the number of automatic interpretation results for many buried objects has increased dramatically.
On the other hand, technology to explain the basis for interpretation results obtained by machine learning is still under development. Moreover, it is assumed that it will take a lot of work to reach a social agreement regarding interpreting such information. Continuing thorough quality control by skilled technicians and supervisors is considered the best option to guarantee the interpretation results. To this end, quality control tools for efficient interpretation are essential. Based on the observation that visual inspection by skilled technicians primarily takes into account the spatial distribution of radar return amplitude anomalies, we propose a tool to aid in interpreting return anomalies in buried objects. Last year, we segmented the interpretation results of the series based on the statistically based value of each pixel based on the change in intensity, and it became intuitively clear that the basis for the proposal could be better expressed as a three-dimensional box.
However, GPR events were only segmented in orthogonal directions and even only in combinations of 2D cross-sections. This year, It is improved to allow cluster segmentation of reflective sections with more realistic slopes.
Interpretation support tools will help build Human-In-The-Loop machine learning systems.
On the other hand, technology to explain the basis for interpretation results obtained by machine learning is still under development. Moreover, it is assumed that it will take a lot of work to reach a social agreement regarding interpreting such information. Continuing thorough quality control by skilled technicians and supervisors is considered the best option to guarantee the interpretation results. To this end, quality control tools for efficient interpretation are essential. Based on the observation that visual inspection by skilled technicians primarily takes into account the spatial distribution of radar return amplitude anomalies, we propose a tool to aid in interpreting return anomalies in buried objects. Last year, we segmented the interpretation results of the series based on the statistically based value of each pixel based on the change in intensity, and it became intuitively clear that the basis for the proposal could be better expressed as a three-dimensional box.
However, GPR events were only segmented in orthogonal directions and even only in combinations of 2D cross-sections. This year, It is improved to allow cluster segmentation of reflective sections with more realistic slopes.
Interpretation support tools will help build Human-In-The-Loop machine learning systems.