Automated AI image recognition tools improve the efficiency of aerial wildlife counts: A multi-species case study on breeding seabirds and pinnipeds at the sub-Antarctic Bounty Islands.

This study demonstrates that combining piloted aircraft surveys with automated AI image recognition significantly improves the efficiency and accuracy of monitoring breeding seabird and pinniped populations at the remote Bounty Islands, reducing processing time from 66 hours to four minutes while maintaining high detection accuracy.

The Gist

Imagine trying to count every single person in a massive, crowded stadium. Now, imagine that stadium is made of jagged rocks in the middle of a stormy ocean, thousands of miles from the nearest city, and the "people" are actually birds and seals that move around, hide in cracks, and look very similar to the rocks they sit on.

That is the challenge conservationists face when trying to count wildlife on the remote Bounty Islands near New Zealand. This paper is about how the researchers swapped a pair of tired human eyes and a clicker for a super-fast, super-smart digital brain to solve this problem.

Here is the story of how they did it, broken down into simple parts:

1. The Old Way: The "Human Marathon"

Traditionally, to count these animals, a team of scientists would have to:

• Fly a plane over the islands.
• Take thousands of high-resolution photos.
• Stitch those photos together into giant maps (like a massive jigsaw puzzle).
• The hard part: A human would then sit at a computer for 66 hours (almost two full work weeks), zooming in on every single pixel, finding the birds, and clicking a mouse to count them one by one.

It was slow, expensive, and the humans would get tired, making mistakes. It was like trying to find a needle in a haystack while wearing thick winter gloves.

2. The New Way: The "Digital Super-Scanner"

The researchers decided to teach a computer to do the counting. They used a type of Artificial Intelligence (AI) called a Deep Learning Neural Network.

Think of this AI like a very hungry, very fast librarian.

• Training: First, they showed the AI about 1,400 pictures of Salvin's albatrosses (the main bird they wanted to count) and told it, "This is an albatross." They also showed it a few pictures of penguins, seals, and other birds, saying, "These are the other guys."
• The Test: Once the AI had "studied" these examples, they gave it the giant photo maps from the flight.

3. The Results: Speed vs. Accuracy

The results were like comparing a snail to a race car:

• Speed: The human took 66 hours to count the birds. The AI took 4 minutes and 6 seconds. That is 963 times faster. It's the difference between walking across a country and flying there in a jet.
• Accuracy: The AI got it right 97% of the time compared to the human expert. It was so good that it could spot the birds even when they were crowded together.
• Bonus Features: Because the AI was smart, it didn't just count the albatrosses. While it was doing its job, it also spotted and counted the penguins, seals, and other birds in the same photos. It was like hiring a security guard who not only counts the people entering the building but also counts the delivery trucks and the stray cats at the same time.

4. Why This Matters

The Bounty Islands are a critical home for the Salvin's albatross, a bird that is "Vulnerable" (at risk of disappearing). To protect them, we need to know how many there are.

• No Disturbance: Sending humans onto these tiny, rocky islands can scare the animals and damage their fragile homes. Using a plane and AI means the animals barely know anyone was there.
• Scalability: If we need to count birds on five different islands next year, we don't need to hire five teams of tired humans for months. We just fly a plane once, and the AI does the rest in an afternoon.
• Future Proofing: The researchers found that if you show the AI about 1,000 examples of a new animal, it learns to recognize it very well. This means in the future, they can easily teach this same tool to count whales, different types of birds, or even ships at sea.

The Bottom Line

This study is a victory for conservation. It proves that we don't have to choose between being accurate and being fast anymore. By combining a piloted airplane (to get the photos) with a smart AI (to count them), scientists can now monitor the health of these remote wildlife populations quickly, cheaply, and without disturbing the animals they are trying to save.

It's like upgrading from counting sheep by hand in a field to using a drone that counts the whole flock in a blink of an eye.

Technical Summary

1. Problem Statement

Accurate population monitoring of seabirds, particularly vulnerable species like Salvin's albatross, is critical for conservation but faces significant logistical hurdles.

• Logistical Challenges: Traditional ground-based surveys in remote sub-Antarctic environments (e.g., the Bounty Islands, New Zealand) are time-consuming, costly, and often impossible due to difficult terrain, weather, and biosecurity constraints.
• Limitations of Existing Aerial Methods: While aerial imagery (from satellites or UAVs) offers a solution, manual counting of animals from these images is prone to human error and is extremely labor-intensive.
• Platform Constraints: Previous AI-based studies have relied on satellite imagery (often too low-resolution for individual bird identification) or UAVs (limited by flight endurance, payload capacity, and camera resolution). There is a gap in applying advanced AI to high-resolution imagery captured by piloted fixed-wing aircraft, which can cover large, remote areas rapidly with superior image quality.

2. Methodology

The study evaluated an automated AI workflow to count wildlife from high-resolution aerial imagery collected via a piloted aircraft.

• Data Acquisition:
  • Platform: Twin turboprop Reims F406 aircraft.
  • Location: Bounty Islands, New Zealand (47°45′S 179°03′E).
  • Imagery: Captured at 1,200 ft AGL using a Nikon D800 full-frame DSLR with a 70-200mm lens (set at 70mm).
  • Processing: 450 individual photographs were stitched into nine ortho-mosaic images covering the entire archipelago.
• Baseline (Human Count):
  • An experienced observer manually counted Salvin's albatrosses from the ortho-mosaics using Photoshop and Mousecount software.
  • Time Investment: Approximately 66 hours (1.5 standard work weeks).
  • Total Count: 30,899 albatrosses.
• AI Model Architecture:
  • Algorithm: Faster R-CNN (Region-based Convolutional Neural Network) with a Residual Network (ResNet) backbone and Feature Pyramid Network (FPN).
  • Inference Strategy: Slicing-Aided Hyper-Inference (SAHI) was employed to handle large images with high object density, improving detection in crowded scenes.
  • Training Data:
    • Primary Target: Salvin's albatross (~1,392 training examples).
    • Secondary Targets: Erect-crested penguins, New Zealand fur seals, and fulmar prions (petrels). These received opportunistic training with fewer examples (ranging from 22 to 1,076) due to lower abundance in training sets.
  • Validation: The model was tested on a subset of images (n=1,410 animals) using a confusion matrix and Intersection over Union (IoU > 0.5) to calculate Precision, Recall, and F1 scores.

3. Key Results

• Processing Speed:
  • The AI processed the re-count of approximately 31,718 albatrosses in 246 seconds (4.1 minutes).
  • This represents a 963-fold increase in speed compared to the 66 hours required for manual counting.
• Accuracy (Salvin's Albatross):
  • F1 Score: 92.8% for the primary target species.
  • Count Comparison: The AI count (31,718) differed from the human count (30,899) by 819 birds (approx. 2.6% difference).
  • Site-Specific Accuracy: Accuracy ranged from 63% to ~100% across different islands, with an overall accuracy of 97% against the manual baseline.
• Multi-Species Performance:
  • Penguins: F1 Score of 89.4%. High accuracy attributed to the high-resolution imagery allowing detection of yellow head crests, despite challenges with shadows and crevices.
  • Seals: F1 Score of 48.0% (limited by low training data: 55 examples).
  • Petrels: F1 Score of 30.0% (limited by low training data: 22 examples).
• Training Data Correlation:
  • An exponential relationship was observed between training data volume and model accuracy. The study suggests a minimum of 1,000 training examples per species is required to achieve >75% model success (F1 score).

4. Key Contributions

• Validation of Piloted Aircraft + AI: Demonstrated that combining high-resolution imagery from piloted fixed-wing aircraft with deep learning is a superior alternative to UAVs and satellites for large-scale, remote wildlife surveys.
• Multi-Species Scalability: Proved that a single model can simultaneously detect and count multiple species (birds and pinnipeds) in mixed breeding colonies, offering significant efficiency gains for future multi-species monitoring.
• Operational Efficiency: Reduced the data processing bottleneck from weeks to minutes, enabling rapid conservation decision-making.
• Training Thresholds: Established empirical guidelines (approx. 1,000 examples) for the training data volume required to achieve reliable detection accuracy for new species in similar environments.

5. Significance and Discussion

• Conservation Impact: The method provides a scalable, non-invasive, and rapid tool for monitoring threatened species (like the Vulnerable Salvin's albatross) in logistically constrained regions. It minimizes disturbance to wildlife compared to ground teams.
• Refinement Potential: The slight discrepancy in counts (AI counting slightly higher) was attributed to the AI counting flying birds, whereas humans counted only nesting birds. Future iterations can be trained to distinguish between nesting and flying individuals to refine population estimates.
• Future Applications: The framework is adaptable. With additional training data, the model can be expanded to count other species, monitor ship-seabird interactions, and track terrestrial mammals, making it a versatile tool for ecosystem management.
• Cost-Benefit: While the initial flight requires a piloted aircraft, the elimination of weeks-long field deployments and the reduction of manual counting labor make this approach highly cost-effective for long-term monitoring programs.

Conclusion: The study successfully demonstrates that integrating AI-driven image analysis with traditional piloted aerial surveys creates a robust, rapid, and accurate methodology for wildlife population censusing, overcoming the limitations of manual counting and smaller-scale drone surveys.