Spatially masked residual learning with a kriging prior for airborne gamma-ray spectrometric interpolation

摘要: We evaluate whether spatially masked graph learning can improve a local ordinary-kriging prior for processed airborne gamma-ray spectrometric responses. The prior is recomputed from visible observations after each training mask, and fitted preprocessing uses only core-training data. Across 397,392 measurements from 108 flight lines, five-fold spatial block cross-validation gives mean R² values of 0.207, 0.607, 0.310 and 0.557 for equivalent uranium (eU), equivalent thorium (eTh), potassium (K) and the total-count response (TC), respectively. The matched prior gives 0.160, 0.603, 0.282 and 0.550. Mean paired gains are therefore +0.047, +0.004, +0.028 and +0.007; eU, eTh and K improve in every fold, whereas TC is slightly worse in one. A paired multilayer perceptron remains near the prior. Heteroscedastic interval calibration yields mean coverage of 0.908-0.913 at a nominal level of 0.90, with increased interval width. An independent coordinate decoder provides arbitrary-coordinate queries at approximately prior-level accuracy. Under reconstructed cross-track-band holdout, eU R² remains negative and gains vary non-monotonically with thinning. The evidence supports a modest, geometry-dependent correction to this prior. It does not establish improved physical concentration accuracy, anomaly detection or transfer to another survey.