We address here the interpolation of missing data in high-resolution geophysical fields under the assumption that a simultaneous low-resolution observation is available. Such multi-scale interpolation is expected to depict geophysically consistent features, especially: i) consistent high-resolution textured patterns, ii) non-Gaussian marginals and iii) specific spectral density. We propose a novel model to address this multi-scale interpolation jointly accounting for the above-mentioned constrains. The key idea is to exploit a non-local or patch-based framework. Such exemplar-based representation provides an implicit texture model, which directly exploits available observations to reconstruct consistent geophysical patterns.