fruitpunch

Guide · AI & data

AI for Wildlife Conservation: Camera Traps, Acoustics and Anti-Poaching

By the FruitPunch research deskUpdated November 15, 20268 min read

In short

AI for wildlife conservation is mostly about sorting sensor data at scale: a 2018 PNAS study automated 99.3% of the 3.2 million Snapshot Serengeti images at volunteer level accuracy, Wildlife Insights trains on over 35 million camera trap images covering 1,295 species, and BirdNET recognizes over 11,000 species by sound. On the protection side, SMART is used at 1,200 sites in over 100 countries and EarthRanger supports 900+ conservation sites, both free to use.

This address once hosted FruitPunch AI, an Eindhoven AI for good community. FruitPunch is a new, independent publication and is not that organization. This guide was written from public sources; the history page explains the change.

AI for wildlife conservation is mainly a way to process the flood of images, sounds and GPS points that field sensors now produce, so that people can spend their time on decisions instead of labeling. The clearest proof is a 2018 study in PNAS showing that deep learning "can automate animal identification for 99.3% of the 3.2 million-image Snapshot Serengeti dataset while performing at the same 96.6% accuracy of crowdsourced teams of human volunteers", saving more than 8.4 years of human labeling effort. Since then the same idea has spread to sound, drones, individual animal identification and ranger patrol software.

Key facts

  • The Wildlife Insights model is trained on over 35 million images across 1,295 species, and blank images can make up 80% of a dataset (Wildlife Insights).
  • Google's SpeciesNet classifier was trained on over 65M images and covers more than 2000 labels (google/cameratrapai).
  • BirdNET recognizes over 11,000 species by sound with open source tools (BirdNET).
  • SMART has been used at 1,200 sites in over 100 countries (SMART Partnership); EarthRanger lists 900+ supported conservation sites and 23K animals tracked by GPS (EarthRanger).
  • Wild Me reports automated identification of individual animals for 324 species (Wild Me).
  • Africa held 22,540 rhinos at the end of 2024, and the poaching rate fell to 2.15%, its lowest level since 2011 (IUCN).

How does AI classify camera trap images?

Camera traps fire on motion and heat, so wind, grass and sun produce huge numbers of empty frames. Wildlife Insights notes that blank images can be up to 80% of a dataset and can take an expert several hours to sort. Most pipelines today work in two steps:

  1. Detection. A general detector finds animals, people and vehicles and draws boxes around them. MegaDetector, originally developed at Microsoft and now maintained as an open source project, is the common choice. Its repository says the model "is trained on several million images from a variety of ecosystems". It does not name species, which is why it transfers well between regions.
  2. Classification. A second model names the species inside each box. Google's SpeciesNet runs as a cascade with MegaDetector, and the repository states that this approach is used for species recognition in the Wildlife Insights platform.

The reason for the split is explained in an arXiv paper from the KDD 2019 Data Mining and AI for Conservation workshop by Sara Beery, Dan Morris and Siyu Yang: models trained in one region can see accuracy fall "catastrophically" in a new region, so a pre-trained general animal detector plus a smaller set of local labels is the more reliable route.

Accuracy depends on how common a species is in the training data. Wildlife Insights says well represented classes like jaguars, mule deer or African elephants generally have between 80% and 99% probability of being correctly predicted, and warns that species with fewer than 1,000 training images may not reach its confidence threshold. In practice, teams let the model remove blanks and label the easy, common animals, then review rare species by hand.

Most users never touch Python. The MegaDetector documentation says AddaxAI is how 95% of users run MegaDetector, and Zamba Cloud, from DrivenData, classifies camera trap videos and images through a web upload and can train a custom model on a user's own labeled footage.

What can AI hear? Bioacoustics and BirdNET

Sound covers ground that cameras cannot: dense forest, night, and animals that are heard far more often than seen. BirdNET, a collaboration between the Cornell K. Lisa Yang Center for Conservation Bioacoustics and Chemnitz University of Technology, uses deep learning "to recognize over 11,000 species globally". It ships as a command line analyzer for large recorder deployments, a phone app, and lightweight models for low power devices such as the BirdNET-Pi.

Acoustic monitoring is also used for protection. Rainforest Connection builds solar powered "Guardian" devices that it says can detect threats in real time, "including but not limited to illegal logging, poaching, blast fishing and forest fires". Its Guardian 3 uses three microphones that capture ambient sound "within a radius of 50 m to 1,500 m depending on the source of the sound".

How is AI used against poaching?

The most widely used anti-poaching tools are not flashy computer vision systems but data platforms that turn ranger patrols, sensor feeds and collar data into maps and alerts.

The pressure these tools respond to is still real. IUCN and TRAFFIC data show that 516 poaching incidents were recorded in Africa in 2024, even as the rhino poaching rate fell to 2.15%. Software does not stop poachers alone; it helps small ranger teams put patrols where the evidence points.

Do drones count wildlife better than people?

Drones are useful where animals gather in large numbers, such as seabird colonies, seal haul outs and herds on open plains. In an experiment with life-sized replica seabird colonies of known size, Hodgson and colleagues found that drone-derived counts were between 43% and 96% more accurate than the traditional ground based counting method, and that counting from the imagery could be semi-automated with a high degree of accuracy.

Thermal cameras on drones extend surveys to night time, when poaching often happens. The Conservation Drones dataset on LILA BC was published for exactly this problem: detecting people and animals in thermal infrared drone video of protected areas. FruitPunch AI, the Eindhoven community that previously used this domain, described its own AI for Wildlife challenge as an effort to build an autonomous UAV based solution for monitoring wildlife reserves. The current FruitPunch site is a separate guides site and is not affiliated with that organization.

Can AI recognize individual animals?

Many species carry natural markings that work like fingerprints: zebra stripes, whale shark spots, giraffe patterns, the trailing edge of a humpback whale's tail. Matching them by eye across thousands of photos is slow, and missed matches bias population estimates.

  • Wild Me, now part of Conservation X Labs, runs the open source Wildbook platforms and reports automated identification for 324 species and 144 publications that use its technology.
  • Happywhale matches humpback whale tail photos from researchers and the public. In a paper in Mammalian Biology, its matching system cut image management time "by at least 98%" and reduced the rate of missed potential matches from about 6 to 9% down to 1 to 3% compared with manual matching.

Individual ID turns casual photos into capture and recapture data, which is how researchers estimate population size, survival and movement.

Which named projects and open tools can people use?

Tool or project What it does Published scale Cost and license
MegaDetector Finds animals, people and vehicles in camera trap images Trained on several million images Free, MIT license
SpeciesNet Species classifier that runs after MegaDetector 65M+ training images, 2000+ labels Free, open source
Wildlife Insights Upload, classify and share camera trap data 35 million images, 1,295 species Web platform
Zamba Cloud Classifies camera trap videos and images, trains custom models Several pretrained models Web platform, open source library
BirdNET Identifies species from sound recordings 11,000+ species Free, open source
Rainforest Connection Acoustic sensors for threats and species Guardian range 50 m to 1,500 m Nonprofit program
SMART Ranger patrol data and reporting 1,200 sites, 100+ countries Free
EarthRanger Real time map of rangers, collars and sensors 900+ sites, 23K tracked animals Free
Wild Me / Wildbook Photo identification of individual animals 324 species Open source

For training and testing models, the LILA BC library (Labeled Information Library of Alexandria: Biology and Conservation) hosts free labeled datasets. Examples include the Snapshot Safari 2024 Expansion with 4,029,374 images from 15 camera trapping projects, about 1.5 million camera trap images from Idaho, 99,909 images from camera traps in western Oregon, and the Big Bird drone set with 49,990 annotations. These are the datasets most published camera trap models are benchmarked against.

What are the limits of AI in conservation?

  • Rare species are the hardest. Classifiers are weakest exactly where conservation interest is highest, because rare animals have few training photos. Wildlife Insights says it can expand to new and rare species when users add data.
  • Models drift between regions. The Beery, Morris and Yang paper shows why local validation is needed before trusting a model at a new site.
  • Connectivity and power. Many reserves have weak networks, which is why BirdNET offers embedded models and Rainforest Connection uses solar powered devices.
  • People make the outcome. Alerts help only if rangers can respond. The IUCN notes that the decline in African white rhinos in 2024 came from poaching plus other pressures such as drought, problems that no detection model fixes alone.

A practical starting point for a field team is to run MegaDetector through AddaxAI on an existing archive of photos, measure how many blanks it removes, and check its errors on a sample before changing any workflow.

Frequently asked questions

What is MegaDetector?

MegaDetector is a free, open source model that finds animals, people and vehicles in camera trap images without naming species. Its repository says it was trained on several million images from many ecosystems, and it is usually paired with a species classifier such as SpeciesNet.

How accurate is AI at identifying animals in camera trap photos?

For common species it can match trained volunteers. The PNAS Snapshot Serengeti study reached 96.6% accuracy on 99.3% of images, and Wildlife Insights reports 80% to 99% correct prediction for well represented classes. Rare species need human review.

Is BirdNET free to use?

Yes. BirdNET offers open source models and a free app, and it recognizes over 11,000 species. It is run by the Cornell K. Lisa Yang Center for Conservation Bioacoustics and Chemnitz University of Technology.

What software do rangers use to fight poaching?

The two most common platforms are SMART, used at 1,200 sites in over 100 countries, and EarthRanger, which supports 900+ conservation sites. Both are free and now cooperate under the SERCA alliance.

Where can I find wildlife datasets to train a model?

LILA BC hosts free labeled camera trap, drone and aerial datasets, including the 4 million image Snapshot Safari 2024 Expansion. Wildlife Insights also has an Explore Data section for shared camera trap data.

Is this site the FruitPunch AI community?

No. FruitPunch AI was an Eindhoven based AI for Good community that ran challenges, including one on drones for wildlife reserve monitoring. This site is an independent guides site that now uses the domain and is not affiliated with it.

Spotted an error in this guide? Email editorial@fruitpunch.ai

Keep reading