Nature intelligence deserves its own playground.

Classification and detection across 12 datasets of birds, land and marine mammals, anurans and insects. By Earth Species Project.

Extends BEANS with call types, life stages, captioning and counting to test zero-shot transfer. The reference eval for audio-language models.

6,800+ training hours over ~10,000 classes, plus eight strongly labeled evaluation sets. ICLR 2025.

Retrieve bird calls from passive recordings after training on citizen-science audio. A hard test of distribution shift.

Annual competition to detect and classify animal calls in soundscapes from autonomous recorders.

Six classification and six segmentation tasks with aggregated reporting and 20 baselines. NeurIPS 2023.

11 datasets of dense tasks across optical, SAR and multi-temporal data. Finds foundation models don't reliably beat supervised ones.

5M+ occurrences of European flora paired with satellite imagery, climate, soil and land-cover data.

Predict bird encounter rates from satellite images using eBird labels, with a Kenya set for low-data regimes. NeurIPS 2023.

Camera-trap photos of 182 species, where each trap is its own domain. Tests generalization to new deployments.

~12K images of 400 IUCN-threatened species held out from training to test zero-shot recognition.

250 expert retrieval queries over 5M iNaturalist images. The best models stay below 50 mAP@50. NeurIPS 2024.