Research map / Vision-language and multimodal / CLIP and zero-shot
Combining Observational Data and Language for Species Range Estimation
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.
Proposes LE-SINR, which embeds iNaturalist presence observations and Wikipedia species text into a shared space to enable zero-shot species range maps from text and text-regularized few-shot estimation.
Problem
Species range map estimation is hard for species with few or no observations and without environmental covariates; existing deep SDMs do not address zero-/few-shot settings.
Approach
Align a location encoder with a text-based species encoder (frozen GritLM embeddings plus an MLP) trained on presence-only iNaturalist data, so free-form habitat and range text yields a species range map without observations; text embeddings also serve as priors for logistic regression in few-shot settings.
Key results
- Zero-shot LE-SINR with range text reaches MAP 0.47 on IUCN and 0.61 on S&T without observations, far above model-mean baseline (0.09 and 0.35).
- Range text priors reduce observations needed for few-shot estimation by about 3x on IUCN and 10x on S&T.
- With environmental covariates, zero-shot range text reaches 0.53 IUCN and 0.64 S&T MAP, matching oracle SINR on S&T scale.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use range descriptions instead of habitat descriptions (ablation) | LE-SINR zero-shot range estimation · IUCN | MAP: 0.28 -> 0.47 | Helped |
| Use range descriptions instead of habitat descriptions (ablation) | LE-SINR zero-shot range estimation · S&T | MAP: 0.51 -> 0.61 | Helped |
| Add environmental features (ablation) | LE-SINR with habitat text · IUCN | MAP: 0.28 -> 0.32 | Helped |
| Add environmental features (ablation) | LE-SINR with range text · S&T | MAP: 0.61 -> 0.64 | Helped |
| Replace habitat/range descriptions with taxonomic text (ablation) | LE-SINR zero-shot range estimation · IUCN | MAP: 0.53 -> 0.21 | Hurt |
| Replace the position encoder with GeoCLIP (ablation) | LE-SINR zero-shot range estimation · IUCN | MAP: 0.53 -> 0.41 | Hurt |
| Replace the position encoder with spherical harmonics encoding (ablation) | LE-SINR zero-shot range estimation · S&T | MAP: 0.64 -> 0.63 | Hurt |
| Replace AN_full with the SatCLIP contrastive loss (ablation) | LE-SINR zero-shot range estimation · IUCN | MAP: 0.53 -> 0.46 | Hurt |
| Increase the negative-sampling value M to as large as 2,048 (ablation) | LE-SINR training · Zero-shot range estimation | MAP | No clear effect |
| Explicitly encourage species tokens and text embeddings to be close (ablation) | LE-SINR training · Range estimation | performance | Hurt |
| Regularize few-shot logistic regression toward a text-derived species embedding rather than zero | LE-SINR text-driven prior · IUCN and S&T | MAP: 10 to 15 points MAP | Helped |
| Use a range-text prior for few-shot estimation | LE-SINR text-driven prior · IUCN and S&T | observations needed: 3× and 10× | Helped |
Surprising findings
- Zero-shot range text on S&T is worth roughly ten observations, and habitat text alone is worth about two.
- Explicitly tying species tokens to text embeddings was found to hurt performance in preliminary experiments.
Limitations
- Training and evaluation data are biased toward North America, Europe, and Australasia.
- Relies on LLM-generated summaries and Wikipedia text that may carry biases and may not be available for rare species.
- No error bars reported and few-shot baselines beyond logistic regression were not compared.
Experimental setup
Datasets: iNaturalist (SINR dataset) Wikipedia species articles eBird Status and Trends IUCN Red List range maps
Benchmarks: IUCN Red List range maps eBird Status and Trends
Metrics: mean average precision
Baselines: SINR Constant Prediction Model Mean LD-SDM GeoCLIP backbone Spherical Harmonics encoding SatCLIP loss
Builds on: SINR LD-SDM GritLM SatCLIP GeoCLIP
Closest papers by meaning
- Feedforward Few-shot Species Range Estimation (ICML 2025 · similarity 0.91)
- Spatial Clustering of Citizen Science Data Improves Downstream Species Distribution Models (AAAI 2025 · similarity 0.86)
- GeoPlant: Spatial Plant Species Prediction Dataset (NeurIPS 2024 · similarity 0.86)
- DivShift: Exploring Domain-Specific Distribution Shifts in Large-Scale, Volunteer-Collected Biodiversity Datasets (AAAI 2025 · similarity 0.85)
- WildlifeMapper: Aerial Image Analysis for Multi-Species Detection and Identification (CVPR 2024 · similarity 0.84)
- RANGE: Retrieval Augmented Neural Fields for Multi-Resolution Geo-Embeddings (CVPR 2025 · 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.