Where species distribution models fail under occurrence-data contamination: calibration error concentrates at stream-network headwaters

Abstract

Species distribution models (SDMs) increasingly inform conservation and biosecurity decisions in freshwater systems, where the reliability of its uncertainty estimates matters as much as its point predictions. Ensemble SDMs derive prediction intervals from across-replicate variance, but this variance captures systematic error only when replicates disagree about it, an assumption that fails when training data are contaminated with low-accuracy records, the norm in citizen-science datasets. Whether this failure is spatially uniform or concentrates in identifiable parts of a range is unknown. Using a panel of European freshwater crayfish spanning native headwater-associated species and invasive lowland colonizers, we show that contamination-induced calibration failure is strongly spatially structured: it concentrates at stream-network headwaters, the topological tops of the network, where upstream-aggregated predictors are structurally undefined, and scales with contamination severity, replicated across four species and both dominant ensemble protocols (replicate and consensus). The failure is driven by upward prediction bias, not by intervals failing to widen: contaminated ensembles overpredict suitability in headwaters, and because the bias is shared across ensemble members, the intervals do not flag it. This is a conservation-relevant blind spot, because headwaters are both refugia for threatened native crayfish and front lines for invasion; an SDM that silently overpredicts suitability there misdirects survey and management effort toward the segments where its predictions are least trustworthy. Standard leave-one-basin-out conformal calibration, the recommended panel-wide remedy, repairs marginal coverage but leaves headwaters undercovered, because a single calibration threshold is dominated by the abundant non-headwater segments. A group-conditional (Mondrian) variant, calibrating the two populations separately, restores reliable coverage in both at no extra cost and reallocates width where it is needed. We recommend network-position-stratified calibration as a default for ensemble SDMs in dendritic freshwater systems.