DualSwin-MAD: a multi-magnetic source identification method based on normalized source strength and multi-scale feature learning

Magnetic anomaly target identification using multiple magnetic dipole sources plays a significant role in resource exploration, unexploded ordnance (UXO) detection, and military target recognition. While Normalized Source Strength (NSS)—derived from magnetic vector gradients—effectively mitigates the influence of source depth and magnetization direction, identifying superimposed signals from closely spaced magnetic sources remains challenging. To address this, we propose DualSwin-MAD, a multi-scale feature identification network for pinpointing magnetic anomaly target locations and magnetic moment parameters. The network employs a large-scale branch to partition anomalous regions and a small-scale branch to extract fine-grained local features, enhanced by physical constraints to boost accuracy. Evaluated on a simulated dataset with 10% noise, DualSwin-MAD achieves 99.51% accuracy in magnetic source counting, with remarkably low Mean Absolute Percentage Errors (MAPE) of 0.020% for location and 0.012% for magnetic moment estimation. demonstrating high-precision identification capabilities.