docs/data_card.md

IndoLepAtlas — Data Card

Dataset Summary

IndoLepAtlas is a computer vision dataset of Indian butterflies, moths, and their larval host plants. It captures the diversity of Indian Lepidoptera across ~1,094 species with images sourced from ifoundbutterflies.org.

Inspired by: iNaturalist
Domain: Indian Wildlife & Biodiversity (CSE3292 §3.6)

Source

• Website: ifoundbutterflies.org
• License: Images are publicly available; photographer credits preserved
• Collection method: Automated scraping with respectful rate limiting

Dataset Composition

Butterflies (~967 species)

• Adult/Unknown life stage photographs
• Early stage (caterpillar, pupa) photographs
• Per-image metadata: species, common name, sex, location, date, credit
• Taxonomic hierarchy: Order → Superfamily → Family → Subfamily → Tribe → Genus → Species

Host Plants (~127 species)

• Hero images and gallery photographs
• Per-image metadata: species, family, location, date, credit
• Cross-referenced with butterfly species they host

Geographic Coverage

Images span all major Indian states with concentration in:

• Kerala, Karnataka, Maharashtra (Western Ghats biodiversity hotspot)
• Northeast India (Assam, Arunachal Pradesh, Meghalaya)
• Andaman & Nicobar Islands

Annotation Process

1. Auto-trimming: Image overlays (text labels) removed via fixed-percentage cropping
2. Auto-detection: Bounding boxes generated using Grounding DINO (zero-shot object detection)
3. Classification: Species labels from source website taxonomy
4. Metadata enrichment: OCR extraction of overlay text (location, date, credit, media code)

Known Limitations

• Some species have very few images (long-tailed distribution)
• OCR may fail on low-quality or unusually formatted overlays → null fields
• Bounding boxes are auto-generated, not manually verified
• Geographic bias toward well-surveyed regions (Western Ghats, Northeast)

Ethical Considerations

• No human subjects or personal information
• Source is a public biodiversity documentation platform
• Photographer attribution maintained via OCR extraction
• Dataset does not include sensitive or endangered species location details