Review Papers Claim by Claim

Visual summary — Data strengths & weaknesses for habitat-suitability analysis

Key takeaways (visual-first)

• Data volume is heterogeneous: global compilations (GridScopeRodents) and citizen-science sources (eBird) provide huge sample sizes that enable fine-grained mapping but introduce sampling bias that must be explicitly corrected and documented GridScopeRodents dataset and ISDM practical guide.
• Method diversity matters: SDM choice (MaxEnt, ensemble methods, ISDMs, macroecological constraints) strongly affects inference for community and species responses; multiple studies show contrasting projections depending on modelling framework and bias-control choices Model-method impacts on community projections.
• Reproducibility & provenance are the bottlenecks: best-practice papers deposit large processed datasets and clear code (GridScopeRodents, moose/Dryad), but many MaxEnt/ArcGIS studies report methods without public code or raw processed layers; reproducibility scores vary and lack of code/data links limits independent validation Data deposition exemplar.

Detailed critique & checklist (what to look for or fix in 'Data for habitat suitability analysis')

1. Data provenance and lineage:

   Any compilation used for SDMs must provide (1) raw occurrence IDs (GBIF download keys, eBird project IDs), (2) cleaning rules (duplicate removal, coordinate uncertainty thresholds), (3) thinning/rarefaction parameters, and (4) temporal filters. High-quality examples include GridScopeRodents (explicit GBIF/iNaturalist DOIs + cleaning) and moose Dryad archiving GridScopeRodents provenance and the moose Dryad archive (data + code) Moose Dryad data.

2. Bias characterization, not just correction:

   Large presence-only datasets (eBird, GBIF) can dominate joint-likelihood fits and bias environmental-response estimates unless weighting or reliability adjustments are applied; the ISDM guide shows explicit weighting w and block validation frameworks to quantify this effect — any review that recommends combining datasets must prescribe (and document) weighting strategies and sensitivity analyses ISDM weighting guidance.

3. Spatial/temporal support (change-of-support):

   When combining IPP (continuous-space intensity) with grid-based survey data, change-of-support is nontrivial. The reviewed literature uses explicit change-of-support transforms (Koshkina et al.; ISDM examples). Any data compilation must document spatial resolution, aggregation rules, and temporal matching; missing this causes biased occupancy estimates and flawed connectivity maps Change-of-support in ISDMs.

4. Choice of algorithm & cross-validation:

   MaxEnt remains standard for presence-only (numerous examples in the list: seagrass, clouded leopard, invasive plants), but algorithm choice and tuning (regularization multiplier, feature classes) and spatial-block cross-validation often change inference; include ENMeval/kuenm tuning outputs and spatial-block results as standard deliverables Clouded leopard MaxEnt tuning and ENMeval tuning exemplar.

5. Connectivity depends on resistance construction:

   Corridor maps (Linkage Mapper, Circuitscape) are useful only if resistance surfaces are justified and sensitivity-tested; clouded-leopard and Amur tiger papers explicitly describe AHP/Gnarly/FRAGSTATS choices and recommend landscape-genetics or telemetry validation where possible Connectivity construction and limits and Tiger connectivity caution.

6. Archiving and reproducible outputs:

   Prefer GeoTIFF/NetCDF outputs, documented parameter files, and code notebooks — GridScopeRodents and the moose Dryad archive are examples to emulate; many MaxEnt/ArcGIS-based papers do not provide enough reproducible artifacts and receive lower reproducibility scores as a result Data deposition example and Seagrass data+limitations.

Where the reviewed "Data for habitat suitability analysis" likely falls short (diagnostic checklist)

• Missing unique identifiers for occurrence records (GBIF download keys, eBird dataset IDs) — prevents exact replication.
• Unspecified cleaning thresholds (coordinate uncertainty, provenance filtering), unclear rarefaction thinning radius.
• No systematic sensitivity test across weighting (w) in integrated models — the ISDM literature shows this change can flip which data source dominates ISDM w-sensitivity.
• Lack of archived model parameter files and raster outputs (e.g., GeoTIFF), limiting downstream reanalysis or ensemble recombination.
• Insufficient documentation of spatial-block CV vs random CV — random CV overestimates transferability for spatially structured data (use blockCV).

Concrete, short reproducible checklist to improve the paper/data

1. List exact downloads: GBIF download DOI(s), eBird project IDs, and date ranges.
2. Provide the full cleaning pipeline (code/notebook) and publish it (GitHub + DOI / Zenodo).
3. For integrated datasets: report the ISDM weighting sweep (w grid), and show block-CV curves of AUC / log-likelihood vs w (choose via objective criterion or report trade-off).
4. Archive final raster outputs (GeoTIFFs), parameter files (.rds/.RData), and minimal metadata (CRS, resolution, extent) in a public repository (Figshare/Zenodo/Institutional data archive).
5. Perform sensitivity analyses: background sampling density, thinning radius, and alternative variable sets; report results in supplemental figures/tables.

Short demonstration: Why weighting matters (visual)

Contrasting examples (what to emulate)

• Emulate GridScopeRodents: public deposition of processed rasters, documented ENMeval tuning, spatial rarefaction details — enables re-use and ensemble recomputation GridScopeRodents reproducibility.
• Emulate Moose (Dryad): share training/testing splits, code to reproduce boosted-regression-tree models and CMIP6 projection steps; this aids transferability testing Moose Dryad archive.

Limitations, blindspots and caveats

1. Combining datasets does not remove ecological mismatches: presence-only records may sample tolerances not realized as viable populations (transient observations), so occupancy/survey data remain essential to infer realized suitability Realized vs potential suitability.
2. Projection uncertainty: climate and land-use futures plus dispersal assumptions drive large uncertainty; map users must get ensemble projections and uncertainty maps, not single-scenario point maps Ensemble climate projections example.
3. Scaling mismatch: marine vs terrestrial datasets require different predictor sets (substrate in seagrasses is critical and missing in many global models), so absence of key predictors should be highlighted prominently Seagrass substrate omission.

Practical next steps for the authors (minimal reproducible deliverables)

1. Publish a data DOI with (a) cleaned occurrence tables with source DOIs/IDs, (b) model parameter files, and (c) final GeoTIFFs for each species/ensemble.
2. Include an R/Python notebook that reproduces key figures and the ISDM weighting sweep, plus a short README describing hardware/seed used.
3. Provide blockCV cross-validation outputs and prediction-transfer diagnostics (MESS/novel predictor maps) so users can judge extrapolation risk.