Ohmic dissipation in conductive media considerably limits the penetrat
ive power of high-frequency electromagnetic imaging methods and implie
s that deep regions can be probed only with low-frequency fields. Unfo
rtunately, these low-frequency fields are governed by a diffusive equa
tion which prevents direct high-resolution imaging as in seismic and g
eoradar imaging. However, a clue for high-resolution imaging in the di
ffusive approximation is given by a Fredholm integral equation of the
first kind which links diffusive fields to their propagative duals. If
these duals could be recovered by inverting this integral equation, t
he seismic imaging toolbox might be used, at least from a theoretical
point of view, to produce fine electromagnetic images. Spectral decomp
osition of the integral operator shows that the inverse problem is num
erically ill-posed for both noisy and/or incomplete data. High-resolut
ion can be achieved only by adding sparsity constraints upon the sough
t solution to the information content of the data. This type of a prio
ri information also strongly regularizes the inversion but implies tha
t the inverse problem must be treated as non-linear. A numerical algor
ithm, designed to work in a continuous parameter space, couples both t
he stimulated annealing and the simplex to recover the propagative fie
ld. Numerical applications for pseudo-data with additive noise reveal
that reflective interfaces can be imaged even within the poorly-favour
able magnetotelluric setup.