Research map / Probabilistic and scientific modeling / Molecules and proteins
Feedforward Few-shot 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 FS-SINR, a Transformer-based feedforward model that predicts a species' range from a few observed locations plus optional text/image metadata, without retraining at inference.
Problem
Estimating spatial ranges of species from very few observations (few-shot), where existing SINR-style methods need per-species retraining and perform poorly with limited data.
Approach
Replace the per-species classifier weights of SINR with a Transformer that takes an unordered set of context location tokens (plus optional text and image tokens) and outputs a species embedding from the CLS token, which is inner-producted with query location embeddings; trained with a batch-restricted asymmetric negative loss so new species need only a forward pass.
Key results
- On IUCN with 10 context locations and text, FS-SINR reaches roughly 0.66-0.72 MAP versus about 0.62 for LE-SINR (text), and it outperforms SINR and other baselines in the few-shot regime.
- Zero-shot with range text, FS-SINR achieves 0.55 MAP on IUCN and 0.67 on S&T (with test species seen in training), and 0.52 and 0.64 without them, versus LE-SINR's 0.48 and 0.60 on the same inputs.
- FS-SINR has 8.2M learnable parameters versus 11.9M for SINR and produces range estimates in about 2% of the CPU time and 6% of GPU time of LE-SINR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FS-SINR with range text instead of LE-SINR with range text | Few-shot range estimation · IUCN, 10 context locations | MAP: 0.62 -> 0.66 | Helped |
| Use FS-SINR with range text instead of LE-SINR with range text | Few-shot range estimation · S&T, 10 context locations | MAP: 0.69 -> 0.72 | Helped |
| Add range text rather than providing no metadata (ablation) | FS-SINR · IUCN, zero-shot | MAP: 0.05 -> 0.52 | Helped |
| Add range text rather than providing no metadata (ablation) | FS-SINR · S&T, zero-shot | MAP: 0.18 -> 0.64 | Helped |
| Use range text instead of habitat text (ablation) | FS-SINR · IUCN, zero-shot | MAP: 0.33 -> 0.52 | Helped |
| Use range text instead of habitat text (ablation) | FS-SINR · S&T, zero-shot | MAP: 0.53 -> 0.64 | Helped |
| Add image metadata instead of using no metadata (ablation) | FS-SINR · IUCN, zero-shot | MAP: 0.05 -> 0.19 | Helped |
| Add image metadata instead of using no metadata (ablation) | FS-SINR · S&T, zero-shot | MAP: 0.18 -> 0.38 | Helped |
| Add image metadata to range text (ablation) | FS-SINR · IUCN, zero-shot | MAP: 0.52 -> 0.46 | Hurt |
| Add image metadata to range text (ablation) | FS-SINR · S&T, zero-shot | MAP: 0.64 -> 0.64 | Mixed |
| Average predictions from three different random-seed models (ablation) | FS-SINR · S&T, zero context locations, Range Text | MAP: 0.64 -> 0.66 | Helped |
| Provide taxonomic rank text through species instead of class-level text (ablation) | FS-SINR · IUCN, zero-shot | MAP: 0.05 -> 0.21 | Helped |
Surprising findings
- Image information does not help on average outside the zero-shot case and can hurt performance when more informative text is available.
- Without any location or text input, the model still learns a spatial prior that yields non-trivial predictions on S&T, which is concentrated in North America.
Limitations
- The model is deterministic given context locations and cannot express the many possible range maps consistent with sparse points.
- Trained on presence-only data; cannot incorporate confirmed absences without modification.
- Citizen-science training data carries location, temporal and taxonomic biases, leading to better performance in North America and Europe.
Experimental setup
Datasets: iNaturalist presence-only species observations (35.5M records, 47,375 species) Wikipedia species text descriptions (127K) iNaturalist species images (200K) IUCN range maps S&T range maps
Benchmarks: IUCN S&T
Metrics: mean average precision (MAP) distance-weighted MAP sparsification error AUC area under the random gain
Baselines: SINR LE-SINR Active SINR Prototype SINR LD-SDM Gaussian Process random forest
Builds on: SINR LE-SINR Active SINR AN-full loss
Closest papers by meaning
- Combining Observational Data and Language for Species Range Estimation (NeurIPS 2024 · similarity 0.91)
- 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)
- Spatial Clustering of Citizen Science Data Improves Downstream Species Distribution Models (AAAI 2025 · similarity 0.84)
- Open-Insect: Benchmarking Open-Set Recognition of Novel Species in Biodiversity Monitoring (NeurIPS 2025 · similarity 0.84)
- WildlifeMapper: Aerial Image Analysis for Multi-Species Detection and Identification (CVPR 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.