Bridging Biodiversity Data Standards and Predictive Ecology: Operationalizing Interoperable Workflows for Species Distribution Modeling in West Africa

What this was about

Peter Ugege reported a controlled experiment asking whether standards-informed processing of occurrence records changes species distribution model outcomes for six West African timber species. Starting from 46,941 records, the standards pathway removed 106 records (placeholders, duplicates, a fuzzy and absent records), but after domain restriction and spatial deduplication both pathways converged on identical 36,407 presence units, so no pathway-specific predictive effect was identifiable. Geographic cross-validation gave ROC-AUC of about 0.66-0.75, far below random cross-validation (~0.99), showing random validation overstates transferability.

Why it matters. It gives a rigorous, reproducible test of the downstream effect of data standards on SDMs and a clear warning about optimistic random cross-validation.

In the room

• Three levels of effect were distinguished: record representation, numerical SDM information and predictive inference.
• Pathway A (minimal processing) kept 46,941 records; pathway B (standards-informed fitness-for-use rules) kept 46,835, a 0.23% difference.
• Within the West African domain the difference fell to 70 records, concentrated in three species and 16 locations, all already represented in pathway B.
• After common spatial deduplication both yielded 36,407 presence units and the same 96,407-row modelling matrix.
• Models: logistic regression, random forest and histogram gradient boosting, with strict geographic cross-validation.
• Random CV mean ROC-AUC ~0.99 vs ~0.70 under geographic hold-out (mean difference 0.28).
• No species was best on every robustness dimension, so no composite score was used.

Notable moments