Abstract
<title>Abstract</title> <p>Fusing multiple geoscience data sources is widely assumed to improve mineral prospectivity mapping (MPM), yet the premise that “more sources beat one” is rarely tested. We recast it as a measurable question and contribute a pre-fusion diagnostic framework rather than a new predictor. Using leakage-free, discriminability-weighted late fusion of per-modality expert models as an analysis vehicle, we define two quantities estimable before fusing: a modality-imbalance index s, the spread of the per-modality discriminative power, and the fusion gain Δ, the improvement of fusion over the strongest single modality. On a balanced Australian Lachlan porphyry-Cu dataset (479 samples, five modalities), nested-cross-validation fusion attains ROC-AUC 0.9944 (tied with equal-weight fusion at 0.9945) — the highest of all methods — yet is statistically indistinguishable from a tuned random-forest baseline; this honest result shows the gain is intrinsically small when modalities are both balanced and highly correlated (mean expert-probability correlation 0.82). The central result is a modality-balance perturbation experiment (12 noise levels × 15 seeds = 180 points): s and Δ are strongly negatively correlated (Pearson r = −0.679), and a segmented-regression change-point test confirms an objective slope break (τ = 0.111, permutation p = 0.0001), with the gain crossing zero near s ≈ 0.08 and saturating thereafter as weighted fusion “safely degrades” to the strongest single modality. A radiometric-dominated rare-earth dataset corroborates the direction. We therefore argue that modality balance and complementarity should be estimated before fusion, not assumed under “the more data the better”. Reported thresholds are study-specific empirical estimates rather than transferable constants, and we discuss how spatial autocorrelation bounds the absolute scores.</p>