We propose a quantum-Hamiltonian-learning-based sequential reconstruction framework for dynamic two-dimensional magnetic fields using a local likelihood model derived from a nitrogen-vacancy center spin-1 Hamiltonian. The latent field is not observed directly; instead, local measurements are generated through nitrogen-vacancy spin dynamics governed by local magnetic-field values and a shared dipolar coupling parameter. Sequential Bayesian updates over overlapping scan windows are combined with temporal posterior propagation to reconstruct the evolving field. Numerical experiments on synthetic maze-like magnetic-field sequences show that the proposed method recovers the dominant spatial structure, achieving a root-mean-square error of 7.037×10⁻⁷ T at the final frame. Adaptive diagnostics show decreasing expected information gain and stable local convergence. Fisher-information and leakage diagnostics reveal a sensitivity–leakage tradeoff under long-interrogation controls, and the implemented adaptive policy incorporates leakage-aware control scoring while leaving the likelihood itself unchanged. Combined horizontal and vertical scans improve reconstruction relative to single-direction acquisition. The shared coupling parameter J is only partially identifiable: its posterior becomes narrow but remains frame-dependent and biased. At the final checkpoint, the coupling posterior reaches J_(rm std)=87.0 Hz, close to a finite-time product-state reference benchmark of 73.3 Hz, while remaining 3.35× above a gain-extrapolated ideal-state benchmark. The posterior mean, however, remains biased by 326.9 Hz, indicating that posterior concentration alone does not imply unbiased coupling identifiability.