Candidate SDM model pool for priority woody invasive plant species (Ailanthus altissima, Acer negundo) in the Danubian riparian corridor

Candidate models were generated to support multi-objective, post-hoc selection of SDMs for woody invasive alien plant species (IAS) in the Danubian riparian corridor. Occurrence data were compiled from opportunistic citizen-science observations and harmonised under a target-group background design, then aggregated to a common 1-km grid that served as the modelling unit across the Danube calibration domain and two independent external evaluation domains (Romania and Serbia). Environmental predictors were screened as described in the associated manuscript.

Model calibration used penalised logistic Generalised Additive Models (GAMs) implemented in Python (pyGAM, LogisticGAM). Candidate configurations were sampled via a Latin hypercube sampling (LHS) design over (i) the smoothness/regularisation parameter λ (sampled log-uniformly across orders of magnitude) and (ii) model complexity (integer number of environmental predictors). For each LHS draw, a predictor subset was randomly sampled from the screened predictor pool under these constraints, and a model was fitted on the Danube calibration domain.

Each candidate model was evaluated using (1) spatially blocked cross-validation on the Danube domain (AUC) and (2) external transfer evaluation (AUC) on Romania and Serbia. These metrics support post-hoc selection following a standard MCDA logic: users can change objective weights (e.g., prioritise external transferability vs internal validation, or penalise complexity) and select the candidate that best matches their criteria. The rationale and definition of selection criteria are provided in the manuscript.

Model artefacts are included to enable re-use. The performance table links each candidate to a stored model via the model_tag / model_name identifier. All fitted candidate models are provided in models_MULTI_LHC.zip (1000 models per species), allowing users to reproduce alternative selections and generate maps for their chosen candidate(s). Model-ready predictors for inference are provided in eco_1000m_z.zip as transformed (where applicable) and z-standardised GeoTIFF rasters aligned to a 1-km ETRS89-LAEA (EPSG:3035) grid. AOI.zip contains the AOI polygon (shapefile) used for optional masking. The scaling-parameter CSV included in eco_1000m_z.zip documents how the provided standardised rasters were produced and is not required if users run inference directly with the provided z-standardised rasters.

Associated manuscript
This dataset accompanies the manuscript by Halabuk et al., “From citizen science records to invasive plant indicators: multi-domain evaluation and scenario-based model selection for the Danube corridor”.

Acknowledgement: “Funded by the EU NextGenerationEU through the Recovery and Resilience Plan for Slovakia under the project No. 09I01-03-V04-00005.”