Research map / Time series / Foundation models for time series
GeoPlant: Spatial Plant Species Prediction Dataset
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.
Introduces GeoPlant, a European-scale plant species distribution dataset with 5M presence-only and 90k presence-absence records plus multimodal predictors, and a Kaggle benchmark with ResNet and MLP baselines.
Problem
Species distribution modeling at high resolution and continental scale is hampered by heterogeneous presence-only citizen-science data and limited exhaustive presence-absence surveys, with no standardized multimodal benchmark for plants.
Approach
A large harmonized multimodal SDM dataset combining PO and PA records with environmental rasters, Sentinel-2 images, climate and Landsat time-series cubes, evaluated on spatially held-out PA surveys using sample-averaged F1 and a Kaggle-hosted benchmark with ResNet-6/ResNet-18/MLP and multimodal ensemble baselines.
Key results
- The multimodal ensemble (Climatic + Landsat + Sentinel-2) reached AUC 94.0 and F1s 34.1 on top-25 predictions, improving to 36.2 F1s with top-K estimation.
- Single-modality ResNet-18 on Landsat cubes reached F1s 29.9 and AUC 91.8, while MLP underperformed (F1s 15.8 to 28.4 depending on modality).
- A naive top-25 baseline by district and bio-geographic zone achieved F1s 20.6%, while PO-trained models scored below 9% naive-style and about 15% F1s with ResNet-18.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Top-K estimation to MME (ablation) | MME with climatic + Landsat · GeoPlant PA test | F1s: 33.8 -> 35.9 | Helped |
| Add Top-K estimation to MME (ablation) | MME with climatic + Landsat + Sentinel-2 · GeoPlant PA test | F1s: 34.1 -> 36.2 | Helped |
| Add Sentinel-2 images to climatic + Landsat MME | MME · GeoPlant PA test | F1s: 33.8 -> 34.1 | Helped |
| Use ResNet-6 instead of ResNet-18 | single-modality climatic cubes · GeoPlant PA test | AUC: 90.5 -> 91.8 | Helped |
| Use ResNet-6 instead of ResNet-18 | single-modality Landsat cubes · GeoPlant PA test | F1s: 29.9 -> 30.3 | Helped |
| Use MLP instead of ResNet-6 | single-modality models · GeoPlant PA test, all modalities | F1s: Climatic: 26.2 -> 22.2; Landsat: 30.3 -> 28.4; Sentinel-2: 22.0 -> 15.8 | Hurt |
| Add Top-K estimation and predict fewer species (ablation) | MME · GeoPlant PA test | Recall: 49.3 -> 45.0; 49.7 -> 45.3 | Hurt |
| Add location predictor (ablation) | XGBoost with climatic, soilgrids, and land cover predictors · GeoPlant PA test | F1s: 26.9 -> 28.7 | Helped |
| Add soilgrids predictor (ablation) | XGBoost with location, climatic, and land cover predictors · GeoPlant PA test | F1s: 28.8 -> 28.7 | No clear effect |
| Increase top-k from 20 to 25 (ablation) | ResNet-6 with Landsat cubes · GeoPlant PA test | F1s: 30.27 -> 30.08 | Hurt |
| Increase top-k from 20 to 25 (ablation) | ResNet-6 with Bioclimatic cubes · GeoPlant PA test | F1s: 26.49 -> 26.98 | Helped |
| Train on PO rather than PA data | single-modality ResNet models · GeoPlant PA test, climatic and Landsat cubes | F1s: PA ResNet-18: 26.2 and 29.9; PO ResNet-18: 8.88 and 15.06 | Hurt |
Surprising findings
- Training on presence-only data and testing on presence-absence surveys performs far worse than even naive baselines based on common species.
- The small ResNet-6 matches ResNet-18 on most single-modality metrics at a fraction of the computational cost.
Limitations
- Presence-only data are biased toward accessible areas and common species.
- Presence-absence data cover limited geography (mainly France, Denmark, Switzerland, Czechia) and about half of the European flora.
- Heavy species imbalance and multimodal, multi-resolution integration remain challenging.
Experimental setup
Datasets: GeoPlant PA surveys (EVA) GeoPlant PO records (GBIF) Sentinel-2 images Landsat ARD time series CHELSA climate SoilGrids MODIS land cover Human footprint rasters ASTER GDEM
Benchmarks: GeoPlant PA test set
Metrics: sample-averaged F1 recall AUC precision
Baselines: ResNet-18 MLP XGBoost MaxEnt Naive top-k baseline
Builds on: GeoLifeCLEF Malpolon
Closest papers by meaning
- Feedforward Few-shot Species Range Estimation (ICML 2025 · similarity 0.86)
- 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)
- Spatial Clustering of Citizen Science Data Improves Downstream Species Distribution Models (AAAI 2025 · similarity 0.84)
- BioTrove: A Large Curated Image Dataset Enabling AI for Biodiversity (NeurIPS 2024 · similarity 0.84)
- GreenHyperSpectra: A multi-source hyperspectral dataset for global vegetation trait prediction (NeurIPS 2025 · similarity 0.84)
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.