Best Paper Award

  Landmine contamination is one of the most significant problems facing regions in which wars have been fought. The effects of landmine contamination can still be destructive, even decades after the end of a conflict, since it restricts land use and mobility, thus hindering social, environmental and economic development. Ground-penetrating radar (GPR) is one technique that has been developed to facilitate the mine-clearing process. Tohoku University has deployed a hand-held GPR sensor called ‘ALIS’ in Cambodia and has detected more than 80 landmines. However, we still have to solve many problems before deploying GPR technologies in actual minefields. For example, GPR measurement has to satisfy Nyquist sampling theorem, resulting in long data acquisition times and long processing times. This is a serious problem in actual mine clearance operations.
   In this study, we analyze the utility of applying an emerging technology, compressive sensing (CS), to landmine detection. Compressive sensing is an advanced signal processing technique which enables sparse signal reconstruction from a very small set of measurement data by exploiting its sparsity. In the case of landmine detection, the mines are usually scattered over a very large area. Since a single GPR survey only covers a small fraction of an area, it is likely that the reconstructed image will only contain a small number of landmines, which makes it a sparsity problem that should be able to be solved by CS. However, in this case, CS faces some challenges because the landmine is not exactly a point target and also faces high level clutter from propagation in the medium.
   To address this issue, we propose an effective technique for estimation of targets by means of GPR using model-based compressive sensing. Model-based CS can improve the performance of CS by allowing only some configuration of non-zero components of the signal, and rejecting solutions which violate the signal model that has been defined. Using a small pixel size, the landmine reflection in the image is represented by several pixels grouped in a three-dimensional plane. This block structure can be used in model- based CS processing for imaging a buried landmine. Evaluation using laboratory data and datasets obtained from an actual minefield in Cambodia shows that model-based CS gives better reconstruction of landmine images than conventional CS. Finally, with its ability to directly process random and irregularly sampled spatial data, model-based CS, applied in conjunction with a GPR handheld device, is expected to offer a faster and more efficient landmine clearing process.