1:45 PM - 3:15 PM
[HTT18-P05] A study of an assistance tool for quality control of machine learning interpretation results on ground penetrating radar data
Keywords:ground-penetrating radar, machine Learning, interpretation, automation
Analysis of ground penetrating radar is generally performed by skilled engineers visually interpreting the shape and spatial distribution of the reflected wave due to the difference in the dielectric constant of the underground structure on the travel-time cross-sectional image. In recent years, a large amount of ground-penetrating radar data has been acquired due to the progress of data acquisition system technology, and the interpretation has been automated and labor-saving. Furthermore, in deep learning, which is one of machine learning, many model architectures have been proposed, and the ability to recognize objects has been dramatically improved. The development and implementation of systems that apply these new technologies to the automatic identification of buried objects, including cavities using ground penetrating radar data are progressing, and some are now commercially available. Therefore, it leads that the automated interpretation results of many buried objects will increase.
On the other hand, the technology to explain the basis of machine learning interpretation results is still under development. In order to guarantee the interpretation result, careful quality control by skilled engineers is still necessary. Therefore, a quality control tool that makes an efficient effort is essential. Accordingly, based on the observation that the skilled engineers' visual inspection significantly respects the spatial distribution of the amplitude anomalies of radar reflections, we study the tools that assist the anomalies of reflections due to buried objects to assist interpretation. This interpretation assist tool is based on segmenting a bunch of interpretation results and statistically based values for each pixel based on changes in intensity to clarify the basis of the suggestion intuitively. The expansion in three dimensions is studied from the correspondence in two-dimension presented last year.
The interpretation assist tools should be helpful for building fully automated explainable machine learning systems in the future.
On the other hand, the technology to explain the basis of machine learning interpretation results is still under development. In order to guarantee the interpretation result, careful quality control by skilled engineers is still necessary. Therefore, a quality control tool that makes an efficient effort is essential. Accordingly, based on the observation that the skilled engineers' visual inspection significantly respects the spatial distribution of the amplitude anomalies of radar reflections, we study the tools that assist the anomalies of reflections due to buried objects to assist interpretation. This interpretation assist tool is based on segmenting a bunch of interpretation results and statistically based values for each pixel based on changes in intensity to clarify the basis of the suggestion intuitively. The expansion in three dimensions is studied from the correspondence in two-dimension presented last year.
The interpretation assist tools should be helpful for building fully automated explainable machine learning systems in the future.