Research map / Probabilistic and scientific modeling / Molecules and proteins
Spatial Clustering of Citizen Science Data Improves Downstream Species Distribution Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Compares ten site-construction approaches for post hoc occupancy modeling of citizen science eBird data, finding spatial clustering (clustGeo, DBSC) that uses both geographic and environmental similarity and keeps all points yields better AUC for 31 Oregon bird species.
Problem
Citizen science checklists lack the repeated-visit protocol required by occupancy models, so sites must be constructed post hoc from opportunistic observations; existing site construction methods discard data or ignore environmental similarity.
Approach
Applies machine learning spatial clustering (clustGeo with Bayesian-optimized parameters, and DBSC) to group eBird checklists into occupancy sites using both geographic and environmental feature distances, keeping all observations, and benchmarks these against ten site-construction approaches.
Key results
- Spatial clustering methods (best-clustGeo, BayesOptClustGeo, DBSC) outperform lat-long and eBird-recommended approaches in percentage AUC improvement across 31 species.
- BayesOptClustGeo, which tunes alpha and lambda automatically with Bayesian optimization, achieves the second-highest mean performance after the oracle best-clustGeo, while DBSC needs no tuning and performs comparably.
- Approaches that discard data (2to10, 2to10-sameObs, 1/UL) rank last, performing worse than lat-long.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use spatial clustering algorithms that incorporate environmental and geographic similarity instead of geographic-only site construction | Occupancy models for eBird bird species · Southwest Oregon eBird test observations | AUC | Helped |
| Retain all available observations rather than discard records when constructing sites | Occupancy models for eBird bird species · Southwest Oregon eBird test observations | AUC | Helped |
| Use 2to10 or 2to10-sameObs site construction, which discards observations, instead of lat-long | Occupancy models for eBird bird species · Southwest Oregon eBird test observations | AUC | Hurt |
| Select clustGeo parameters using test performance (best-clustGeo) rather than Bayesian optimization | clustGeo occupancy models · Southwest Oregon eBird test observations | AUC | Helped |
| Vary clustGeo parameters α and λ across parameterizations (ablation) | clustGeo occupancy models · 31 bird species | AUC | Mixed |
Surprising findings
- Single-visit sites (SVS), which ignore the occupancy closure assumption, are a close competitor to the top method in AUC, though the authors caution about parameter identifiability.
Limitations
- Performance is measured only by predictive AUC/AUPRC on held-out observations, not by accuracy of latent occupancy estimates, which cannot be validated directly.
- best-clustGeo uses test-set oracle tuning and is not directly comparable; BayesOptClustGeo shows a performance gap and higher variance.
- Study limited to one region (southwest Oregon), two years, and 31 bird species; occupancy model closure assumed via breeding-season window.
Experimental setup
Datasets: eBird basic dataset (Oregon, 2017-2018) Landsat Tasseled Cap features DEM elevation
Benchmarks: eBird basic dataset held-out 2018 checklists
Metrics: AUC AUPRC percentage AUC improvement over lat-long
Baselines: SVS 1/UL lat-long 2to10 2to10-sameObs rounded-4 1-kmSq best-clustGeo
Builds on: clustGeo density-based spatial clustering (DBSC) occupancy models Bayesian optimization unmarked
Closest papers by meaning
- Combining Observational Data and Language for Species Range Estimation (NeurIPS 2024 · similarity 0.86)
- DivShift: Exploring Domain-Specific Distribution Shifts in Large-Scale, Volunteer-Collected Biodiversity Datasets (AAAI 2025 · similarity 0.85)
- GeoPlant: Spatial Plant Species Prediction Dataset (NeurIPS 2024 · similarity 0.84)
- Feedforward Few-shot Species Range Estimation (ICML 2025 · similarity 0.84)
- WildlifeMapper: Aerial Image Analysis for Multi-Species Detection and Identification (CVPR 2024 · similarity 0.83)
- BirdCollect: A Comprehensive Benchmark for Analyzing Dense Bird Flock Attributes (AAAI 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.