SE-EDSR: A Deep Learning Method for Gas Distribution Mapping

Deep learning methodologies for chemical dispersion map reconstruction

“…In Gas Distribution Mapping the target is to reconstruct a dense representation of the distribution usually from a limited set of sampling points, that could be considered sparse sampling data. From this point of view, as proposed in Zhang et al 73 , GDM can be considered similar to the SISR problems where a low-resolution matrix describes the input, and the final output is to reconstruct the high-resolution image. By using the learning approach, a dense representation is constructed by considering the spatial correlation of the element in the sparse matrix by introducing deep network models.…”

“…In literature, model-based and statistical-based approaches have been widely explored and utilized for various applications, including environmental monitoring, gas distribution mapping, and hazard assessment. These approaches leverage mathematical models or statistical methods to analyse and interpret complex datasets, providing valuable insights into the behaviour and characteristics of environmental phenomena [73][74][75] . An introduction of different methodologies for dispersion mapping is provided in Reggente 75 .…”

“…For example, the model-based methods rely on the prior gas dispersion model, but the turbulent dispersion model cannot be directly used for parameter estimation due to its complexity 76 . The statistical-based methods at the same time suffer from hand-design parameters, such as in the KDM where such parameters have a significant impact on the mapping results 38,73,77 .…”

“…To overcome the dependence of the model-based and statistical-based methods on predefined parameters that influence reconstruction performances, a new approach based on deep learning has been proposed in the paper as the initial study of the applicability of such methodologies. The first attempt to introduce such methodology in the field of GDM was proposed by Zhang et al 73 , starting from the similarity between the process of building dense information from a sparse sampling of GDM and Single Image Super-Resolution (SISR). In general, SISR targets to recover high-resolution images from lowresolution images by using intrinsic relationships between pixels.…”