Improving the spatial resolution of magnetic field measurements is essential for accurately characterizing the distribution of magnetic sources. In Scanning Magnetic Microscopy (SMM), the measured out-of-plane component B z reflects the geometry of the magnetization near the sample surface. However, limitations imposed by sensor-to-sample distance and sensor size introduce spatial blurring, masking fine magnetic structures and reducing the reliability of quantitative estimates. In this study, we propose a hybrid approach that combines downward continuation (DC) techniques with convolutional neural networks (CNNs) to reconstruct high-resolution magnetic field maps from low-resolution experimental data. By numerically projecting the measured field onto virtual planes closer to the sample and applying CNN-based refinement, the method enhances spatial contrast, reveals localized magnetization features, and improves the accuracy of subsequent magnetic moment estimation. The approach was validated using both synthetic and experimental data, including iron oxide particle systems and geological rock slices. Our results demonstrate that the method significantly improves spatial definition without compromising the physical integrity of the measured field, reducing signal losses by up to 86%. • A hybrid downward continuation + CNN framework is introduced to enhance the spatial resolution of Scanning Magnetic Microscopy (SMM). • A physically grounded lift-off estimation using spectral decay allows model-independent reconstruction of the magnetic field. • The methodology achieves up to 77% reduction in magnetization underestimation compared to raw experimental data. • The approach resolves fine-scale magnetic sources previously masked by sensor–sample separation and Hall sensor averaging. • Application to geological thin sections reveals eight magnetization centers, compared to only five in the raw data at higher lift-off.
Sinimbu et al. (2026) studied this question.