Automatic Extraction of Magnetic Lineaments Using U-Net: Application to the Peixoto de Azevedo Area, Alta Floresta Gold Province, Brazil

Magnetic lineaments are usually delineated from derivative-based images through visually selected thresholds, a subjective procedure that is difficult to reproduce over large surveys. We present an automatic workflow based on semantic segmentation. The input attribute is built in two stages: the total gradient of the total magnetic anomaly, which, for dyke-like and contact-like sources, is independent of the magnetization direction, followed by its tilt derivative, which equalizes the response of weak and strong sources within a bounded range. This combination avoids reduction to the pole, which is unstable at the low magnetic latitudes of both surveys used here. A dual-output U-Net, predicting both a lineament mask and its skeleton, was trained on aeromagnetic data from the Carajás Province and applied without retraining to the Peixoto de Azevedo area, in the Alta Floresta Gold Province. The decision threshold was calibrated by maximizing the global Dice coefficient on the validation set, giving t∗=0.335. Applied to a synthetic model of prescribed geometry, the workflow recovered the six bodies of 20 nT or more over 84 to 100 percent of their length, including the one striking almost north–south, whose response is attenuated sevenfold at this latitude; the bodies of 12 and 7 nT were not recovered even in noise-free data, placing the detection limit between 12 and 20 nT. The median positional error is below one grid cell, and the offset normal to the trace stays under 15 m and changes sign among the bodies, so the enhancement introduces no systematic displacement. At the regional scale, the workflow produced 1886 segments dominated by east–west orientations, consistent with the regional structural framework and unchanged when the segments are weighted by their length. The workflow offers a reproducible alternative to manual lineament mapping and delivers GIS-ready structural products for regional mapping and mineral exploration.