Species identification & camera-trap biodiversity monitoring

Knowing which species occur where is the basis of nature conservation, environmental impact assessments and biodiversity reporting, but manually sorting millions of camera-trap images or identifying thousands of look-alike taxa is slow and expert-limited. Fine-grained classifiers and detectors now remove empty frames, propose species labels and let a smartphone name a plant in seconds, so scarce experts only review uncertain, rare or decision-critical records.

Problem & value

• Today: a single camera-trap survey produces hundreds of thousands of images, most of them empty (wind, sun, grass). Ecologists or volunteers tag them by hand. Citizen scientists depend on field guides or forum experts. Agencies and consultancies need defensible species lists for permits.
• AI value: a detector discards empty frames and flags animals, people and vehicles; a regional classifier proposes the species; experts verify only low-confidence or protected-species records. Photo apps give private users and schools instant suggestions and feed verified observations into shared databases.
• Drivers in the EU: the Nature Restoration Regulation (EU) 2024/1991 requires Member States to monitor and restore ecosystems; CSRD/ESRS E4 requires companies to report on biodiversity impacts; wind-farm and infrastructure permitting require species surveys [1][2].
• Reported accuracy depends heavily on region and class mix; Google describes SpeciesNet as covering about 2,500 animal categories trained on 65 million labelled images, and recommends human review of uncertain, rare or management-critical records [3]. No generic speed-up figure is claimed here.

Variants

• Camera-trap batch processing (ecology, forestry, hunting associations): detect, classify, count, export.
• Citizen-science photo identification (private, schools): Pl@ntNet, Flora Incognita, iNaturalist, Seek, Obsidentify; observations flow into GBIF.
• Impact assessment and wind-farm permitting: standardised survey stations, bat and bird focus, auditable species lists; links to VIS-060.
• ESRS E4 corporate reporting: site-level biodiversity baselines for companies with land holdings.
• Smart bird feeder (edge variant): Birdbuddy-style consumer camera that names visiting birds.
• Insects and fungi: moth traps, pollinator monitoring, mushroom ID (high risk of error: never use for edibility decisions).
• Out of scope: eDNA metabarcoding is not vision (see SCI-009); bird song is audio (SPEECH-015).

How it works

1. camera trap / smartphone photo + GPS + timestamp
2. ingest, deduplicate bursts, read EXIF/station metadata
3. detector (animal | person | vehicle | empty) -> crops
4. fine-grained classifier (regional taxon list, top-k + confidence)
5. geographic/seasonal filtering with checklist (occurrence plausibility)
6. human review queue (rare, protected, low-confidence, person frames)
7. counting / event aggregation (independent detections, individuals per burst)
8. export (Darwin Core, GBIF, Wildlife Insights-style platforms) + report

• Two-stage design: MegaDetector finds animals, people and vehicles but not species; a second model (SpeciesNet, DeepFaune, regional or custom) names the species on the crop [3][4].
• Geographic priors: restricting predictions to species plausible for the location and season removes many impossible answers.
• Human-in-the-loop: experts confirm protected, invasive and first-record species; person frames are deleted or blurred before sharing; confidence thresholds are set per taxon group.
• Vision-language models can describe unusual scenes and handle long-tail cases, but are not calibrated taxonomic classifiers and should not replace the specialised models for counts.

Data preparation

Data is camera-trap images and video (often 100,000 to millions of frames), plus GPS, timestamps, station metadata and regional species checklists. In practice they sit on SD cards, laptops, external disks and shared drives, with inconsistent folder names.

Typical gaps: empty or false-trigger frames (often over half), duplicates from burst shots, missing or wrong timestamps, no station and deployment metadata, and no labelled reference images for rare local species. Preparation steps: ingest all cards into one structure with checksums, read EXIF time and location, record deployment sheets per station, run a detector such as MegaDetector to remove empty frames, then classify. Build a verified test set of a few thousand images labelled by an expert to measure accuracy per species. Blur or discard frames showing people, and restrict locations of sensitive species.

Effort share: the largest part of the project, about 40-60 percent. Related groundwork: PREP-006 (digitisation of archives), PREP-010 (metadata and tagging), PREP-018 (evaluation sets), PREP-020 (harvesting labels from existing records) and DATA-003.

Implementation paths

Which path for whom

• public: C - survey projects process very large image sets, and open models such as MegaDetector and SpeciesNet run on modest GPUs with data staying in-house.
• enterprise: C - ecological surveys and land management benefit from a shared pipeline, with predictable cost at high image volumes.
• developer: C - open models and tools can be embedded and fine-tuned for regional species without per-image fees.
• private: A - free apps such as Pl@ntNet or iNaturalist identify single photos with no setup.
• smb: C - ecology consultancies, forestry and landowners can run an open-source pipeline on HW2 and reuse it across projects, which beats per-image fees.

Path A — Buy (off-the-shelf)

• Citizen-science apps (free): Pl@ntNet (France, plants), Flora Incognita (Germany, TU Ilmenau / MPI-BGC, plants), iNaturalist (global, community verification), Obsidentify (NL, Naturalis), Merlin (Cornell, birds). Several are EU-run.
• Camera-trap platforms: Wildlife Insights (Google Earth Outreach and partners), TRAPPER (open source, Poland), Agouti (NL, Wageningen), Timelapse, Zooniverse for crowd labelling; commercial wildlife-camera makers (Reconyx, Browning, Swarovski/Spypoint-type cellular cameras) increasingly bundle on-camera species tags.
• Choose A for one-off surveys or schools; check data location, licence of uploaded images and export formats.

Path B — Build on cloud APIs

• Pl@ntNet API (plant ID, quota-based) and the iNaturalist computer-vision suggestion endpoints for apps.
• General VLMs (Claude, Gemini, GPT, Mistral) for descriptive triage and metadata extraction, always with a species-specific second opinion. Use EU regions for any imagery containing people.
• Storage and compute: object storage in an EU region, batch jobs running the open-source models of Path C.

Path C — Self-host (open source), recommended for survey projects

• Detector: MegaDetector (v5/v1000 family, MIT-style licence) via the megadetector Python package [4].
• Classifiers: SpeciesNet (Google, Apache-2.0, global ensemble) [3]; DeepFaune (French CNRS, European fauna, also GUI), DeepFaune New England for North America [6]; AddaxAI (formerly EcoAssist) and Biowatch as desktop front ends.
• Review and data management: Timelapse, TRAPPER, Camtrap DP data standard for exchange.
• Hardware: HW1 for on-camera or laptop inference with small models; HW2 (one 24 GB GPU) processes millions of images per week.

Path D — Custom training

Worth it for local fauna (e.g. UK mammals, Mediterranean birds) where global models confuse congeners or lump birds into one class [6]. Needs a few hundred verified crops per species, balanced across sites, seasons and day/night (infrared) conditions; always hold out entire camera stations for testing, not random frames.

State of the art (2026)

• Two-stage pipelines are standard: a general detector plus regional classifiers; SpeciesNet was released as open source and combines MegaDetector with an EfficientNetV2 classifier and geographic filtering [3].
• Foundation models: BioCLIP-style vision encoders trained on the Tree of Life dataset enable zero- and few-shot taxon recognition, and open-vocabulary detectors help with unseen classes.
• Limits and failure modes: domain shift to new sites and camera models, night infrared images, partial animals, small mammals and rodents, look-alike species, and rare classes with few examples. Per-class accuracy can differ sharply from overall accuracy. Individual counting of herds and flows of animals remains error-prone.
• Fungi and insects are the weakest groups; photo-only identification of mushrooms must not be used to judge edibility.
• Label noise in public training sets propagates into models; validate on your own stations.

Shared resources & combinations

• Same stack (object-detection + image-classification): VIS-014 livestock monitoring, VIS-005 plant disease, VIS-001 inspection models, VIS-021 satellite analysis for habitat context.
• Edge deployment: Jetson/Hailo cameras shared with VIS-013 wildfire detection and VIS-060 bird-collision systems.
• Downstream: biodiversity indicators for ESRS E4 and nature-restoration reporting; VIS-049 (deforestation-free supply chains) uses species/habitat evidence; VIS-030 anonymisation for person frames; SCI-009 for eDNA complements.
• Data standards: Darwin Core, Camtrap DP, GBIF publishing.

Effort, cost & KPIs

• Path A: days; free apps or platform subscriptions.
• Path B: weeks for an app with API integration; cost per call or quota.
• Path C: 2-6 weeks to a working pipeline on existing hardware; main cost is expert validation time.
• Path D: months, dominated by labelling and per-station validation.
• KPIs: share of empty frames removed (with recall of true animal frames), per-species precision/recall on held-out stations, share of images needing human review, expert hours per 10,000 images, detection-event quality, time from retrieval of SD cards to validated dataset.

People & change

Involve the project lead (department lead) and the ecologists who validate results from the start, since their expertise defines the species list and acceptance threshold. Bring in IT for storage and GPU, data owners such as landowners or authorities who grant survey permission, and management for budget and publication rules.

Works council: usually no, as the system does not monitor employees. It can become relevant if field staff are photographed or tracked by the cameras or if GPS tracks measure their performance (section 87(1) no. 6 BetrVG); then inform the council (section 90(1) no. 3 BetrVG). Public bodies check with the staff council.

Data protection: cameras can capture walkers, hunters or staff, so involve the data protection officer, check the legal basis, apply signage and deletion of human frames, and run a DPIA (Art. 35 GDPR) if cameras cover paths or public areas; sign an AVV (Art. 28 GDPR) with any cloud provider.

AI literacy: support staff and volunteers with training on error types and confidence thresholds (Art. 4 AI Act). Communicate to the public and landowners what is recorded. Pilot with one site and one season. Keep humans in control: experts verify rare, protected or low-confidence species before results enter reports or legal decisions.

Guides: GUIDE-005, GUIDE-009, GUIDE-006.

Risks, regulation & compliance

• EU AI Act: minimal risk for species identification; AI literacy duty (Art. 4) applies. Note Art. 5 does not apply as long as no individual is identified. Do not repurpose camera traps for identifying or tracking people (poaching enforcement or surveillance of visitors raises biometric and policing questions that are outside this use case).
• GDPR: camera traps in forests and parks capture hikers, hunters, workers and licence plates. Treat as personal data: legal basis (public task or legitimate interest), signage, DPIA where cameras cover public paths, automatic deletion or blurring of human frames, retention limits. Location of a private household or a farm can itself be personal data [7].
• Sensitive locations: publishing precise coordinates of threatened or poached species (rare birds, orchids) can enable harm; obscure or generalise coordinates before open publication.
• Quality and liability: permitting reports need documented expert verification; AI output alone is not a survey.
• Safety: plant and mushroom apps give wrong answers; add visible disclaimers for edibility and toxicity.
• Mitigations: human review of protected species, site-held-out validation, versioned models, audit trail of AI vs final label.

Getting started

1. Define the question (presence list, occupancy, abundance) and pick the data standard (Camtrap DP / Darwin Core).
2. Run MegaDetector on a sample of 5,000 of your own images and measure empty-frame removal and animal recall.
3. Add SpeciesNet or DeepFaune, restrict to the regional species list, and compare with expert labels per species.
4. Set up a review queue for low-confidence, protected and person frames, with automatic deletion/blurring of humans.
5. Publish verified records to GBIF or a national portal with coordinates generalised where needed.

Sources

1. Regulation (EU) 2024/1991 on nature restoration — https://eur-lex.europa.eu/eli/reg/2024/1991/oj
2. EFRAG — ESRS E4 Biodiversity and ecosystems (Delegated Regulation (EU) 2023/2772) — https://eur-lex.europa.eu/eli/reg_del/2023/2772/oj
3. Google Research — Where wild things roam: Identifying wildlife with SpeciesNet — https://research.google/blog/where-wild-things-roam-identifying-wildlife-with-speciesnet/
4. MegaDetector documentation — https://github.com/agentmorris/MegaDetector/blob/main/megadetector.md
5. Google cameratrapai (SpeciesNet) repository — https://github.com/google/cameratrapai
6. DeepFaune New England, species classification for trail camera images — https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12617260/
7. Regulation (EU) 2016/679 (GDPR) — https://eur-lex.europa.eu/eli/reg/2016/679/oj
8. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj