Abstract. Paleovalleys are buried ancient river valleys that often form
productive aquifers, especially in the semiarid and arid areas of
Australia. Delineating their extent and hydrostratigraphy is however a
challenging task in groundwater system characterization. This study
developed a methodology based on the deep learning super-resolution
convolutional neural network (SRCNN) approach, to convert electrical
conductivity (EC) estimates from an airborne electromagnetic (AEM) survey in
South Australia to a high-resolution binary paleovalley map. The SRCNN was
trained and tested with a synthetic training dataset, where valleys were
generated from readily available digital elevation model (DEM) data from the
AEM survey area. Electrical conductivities typical of valley sediments were
generated by Archie's law, and subsequently blurred by down-sampling and
bicubic interpolation to represent noise from the AEM survey, inversion and
interpolation. After a model training step, the SRCNN successfully removed
such noise, and reclassified the low-resolution, converted unimodal but
skewed EC values into a high-resolution paleovalley index following a
bimodal distribution. The latter allows us to distinguish valley from
non-valley pixels. Furthermore, a realistic spatial connectivity structure
of the paleovalley was predicted when compared with borehole lithology logs
and a valley bottom flatness indicator. Overall the methodology permitted us to
better constrain the three-dimensional paleovalley geometry from AEM images
that are becoming more widely available for groundwater prospecting.