Neural Networks in Biodiversity Monitoring: Listening and Watching for Species at Risk

The planet is in the middle of what biologists call the sixth mass extinction event — species disappearing at rates estimated to be 100 to 1,000 times higher than natural background extinction rates. Tracking this crisis requires knowing where species are, how their populations are changing, and when illegal activity threatens them. The traditional approach — field biologists counting animals by hand, analyzing scat and tracks, reviewing camera trap images one by one — is too slow and too expensive to match the scale of the problem. Neural networks are beginning to change that equation.

Key takeaways

• Deep learning models can identify species from camera trap images with accuracy exceeding 90%, processing in minutes what would take human reviewers months.
• Bioacoustic AI systems can detect endangered species calls in continuous audio recordings, enabling 24/7 passive monitoring across vast areas.
• Anti-poaching systems combining AI with acoustic sensors have shown real reductions in illegal activity in protected areas where they've been deployed.
• Challenges remain in generalizing models across ecosystems and in building tools that are usable by under-resourced conservation organizations.

The Data Flood Problem in Conservation Science

Camera traps — motion-triggered cameras placed in wildlife habitats — have become ubiquitous in conservation research. A study covering a single national park might deploy hundreds of cameras for a year, generating millions of images. Reviewing those images to identify species, count individuals, and assess animal behavior requires enormous human effort. In many research organizations, graduate students spend months of their time on this task alone — time that could go into analysis and action.

Audio monitoring presents the same challenge at greater scale. Autonomous recording units placed in forests or reefs can record continuously for weeks. A single unit generates thousands of hours of audio. Acoustic ecologists studying bird populations or coral reef health cannot listen to all of it.

Neural networks — particularly convolutional neural networks (CNNs) for images and spectrogram-based models for audio — offer a way to automate the sorting, classification, and counting that bottlenecks conservation science.

Camera Trap Species Identification

The Wildlife Insights project, a collaboration between Google AI, World Wildlife Fund, Wildlife Conservation Society, and several research universities, provides a free cloud platform for camera trap analysis. Their deep learning models, trained on tens of millions of labeled wildlife images from around the world, can classify 614 species and distinguish empty frames from frames with animals. In benchmark tests, the models achieve over 90% accuracy on common species and significantly reduce the human review burden — staff focus on verifying uncertain cases rather than classifying everything from scratch.

A 2021 paper in the Proceedings of the National Academy of Sciences quantified the time savings: AI pre-classification reduced human review time by 75–90% depending on species composition, while maintaining equivalent scientific quality.

Snapshot Serengeti, run by the University of Minnesota, leveraged crowdsourced human classifiers from the Zooniverse citizen science platform in combination with CNN classifiers to produce a multi-year dataset of Serengeti wildlife — one of the largest camera trap datasets ever assembled. The hybrid approach validated AI classifications against consensus human answers, producing high-confidence species counts for lions, cheetahs, wildebeest, and dozens of other species.

Conservation Metrics and similar companies provide AI-powered camera trap analysis services to government agencies and NGOs in Latin America and Southeast Asia, where field staff may lack the expertise or time for manual image review. Critically, these services are beginning to make population trend monitoring feasible in regions where it was previously impossible due to cost.

Bioacoustics: Hearing Ecosystems

Sound is an underused dimension of ecological monitoring. Many threatened species — bats, whales, certain amphibians, migratory birds — are easier to detect acoustically than visually. The Amazon's bird community can be partially assessed by audio recording alone, because many species are rarely seen but regularly heard.

BirdNET, developed at the Cornell Lab of Ornithology in partnership with the Chemnitz University of Technology, is a neural network model that identifies bird species from short audio clips. Available as a free app and API, it can identify over 6,000 species worldwide. A smartphone user recording ambient sound receives real-time species identification. Researchers deploy BirdNET on arrays of recording units to generate continuous bird community data from forests, farms, and wetlands.

Rainforest Connection (RFCx) takes a different angle: using acoustic AI not just for species monitoring but for real-time illegal logging and poaching detection. Their system deploys repurposed smartphones in forest canopies, streaming audio to cloud servers where ML models detect chainsaw sounds, gunshots, and logging truck noise. When detected, the system alerts park rangers with GPS-located alerts. In deployments in Ecuador, Sumatra, and the Congo Basin, RFCx reports a measurable reduction in illegal chainsaw activity in protected areas monitored by their system — though peer-reviewed impact assessments are still limited.

NOAA's Passive Acoustic Monitoring program uses deep learning to analyze hydrophone recordings for endangered whale vocalizations. Identifying right whale calls in hundreds of hours of ocean audio — necessary for enforcing speed restrictions on shipping lanes — was previously impossible to do in real time. ML classifiers now process hydrophone feeds in real time, enabling dynamic routing alerts to ships when whales are present.

Individual Animal Recognition

Beyond species identification, AI is enabling individual animal recognition — essentially facial recognition for wildlife. This matters for population studies because counting individuals is far more informative than counting detections.

Wildbook, developed by Wild Me (now part of Conservation X Labs), uses computer vision to identify individual whales, whale sharks, manta rays, giraffes, and other species from distinctive markings. A whale's fluke notch pattern is as distinctive as a fingerprint. Wildbook matches submitted photos against a growing database, effectively crowdsourcing population surveys from citizen science submissions, research cruises, and tourist boat operators.

For the North Atlantic right whale, one of the world's most endangered large mammals with fewer than 340 individuals remaining, Wildbook's catalog enables researchers to track individual whales' movements, reproductive history, and entanglement injuries over decades — a longitudinal database that would be impossible to maintain without automated matching.

At a very different scale, researchers from the University of Bristol used deep learning to identify individual bumblebees from video frames — allowing behavioral ecologists to track individual foraging trips and map colony activity without physical marking.

Satellite and Drone-Based Species Monitoring

Neural networks applied to satellite imagery are opening new possibilities for monitoring species that are hard to approach on the ground. Imperial College London and partners have used high-resolution satellite imagery combined with object detection models to count penguins in Antarctic colonies, seals on ice floes, and wildebeest herds on the Serengeti — from orbit.

The approach requires training data (manually labeled images where the species of interest has been identified), high enough resolution imagery to distinguish animals from background, and careful model validation in varied lighting and terrain conditions. These constraints limit current application to large animals in relatively open habitats, but satellite resolution continues to improve.

The Allen Institute for AI developed a dataset and benchmark — iNat — from citizen science observations on iNaturalist that has become a standard for training fine-grained species recognition models. These models have now been integrated into iNaturalist's identification system, helping the platform's 4 million users get accurate species IDs for fungi, plants, insects, and animals.

Limitations and Conservation Realities

Rare species present the classic long-tail problem: the species conservation science most needs to monitor are often the ones least represented in training data. Models trained on common species may fail on rare ones, precisely when failure is most consequential.

Model generalizability across ecosystems is imperfect. A CNN trained on Serengeti camera trap images may perform poorly in Amazon cloud forest, where lighting, vegetation density, and species composition are completely different.

Cost and connectivity constrain deployment in the most biodiverse regions. Tropical forests and coral reefs are often in areas with minimal connectivity, requiring local edge computing or periodic data collection rather than continuous cloud analysis.

Conservation impact beyond data collection is the Eventually, test. Better species data is necessary but not sufficient — the biodiversity crisis requires political will, funding, and changes to land use and economic incentives that no algorithm provides.

FAQ

How accurate are neural networks at identifying species in camera trap images? For common, well-represented species, accuracy exceeds 90% in benchmark tests. For rare species with few training examples, accuracy can drop significantly — sometimes below 60%. Practical deployments typically use AI to pre-sort images and flag uncertain cases for human review, rather than operating fully autonomously.

Can AI detect new or undescribed species? AI classification models identify species based on training data — they can't assign a name to a new species. However, anomaly detection approaches can flag images or audio recordings that don't match known categories, potentially flagging candidates for human expert review. Several cases of researchers being alerted to unusual animal captures through AI anomaly flags have been reported informally.

Is bioacoustic monitoring better than camera traps for some species? Yes, significantly. Bats, many frog species, most whale species, and nocturnal forest birds are far more effectively monitored acoustically. The two approaches are complementary and increasingly used together in comprehensive biodiversity monitoring protocols.

How does AI help anti-poaching specifically? AI supports anti-poaching through acoustic detection of gunshots and chainsaw sounds, analysis of patrol data to predict where poaching is most likely (used by tools like PAWS, developed at USC), and rapid processing of camera trap images to flag human presence in protected areas. These tools support rangers rather than replacing them.

What's the biggest limitation of current biodiversity AI? Training data imbalance is arguably the most fundamental. Models perform well on species with thousands of labeled examples and poorly on species with few. Since the species facing the greatest extinction risk are typically rare, the training data problem is also a conservation priority problem. Efforts like the Global Biodiversity Information Facility are working to improve labeled data availability.

---

Read next:

• Neural Networks for Ocean Monitoring — How AI Protects Marine Ecosystems
• AI-Powered Carbon Capture Technology — Can Machine Learning Scale Carbon Removal?
• Countries with the Most Marine Protected Areas (2025 Ranked)