Fine-scale habitat partitioning of sympatric stingrays revealed by drone-based remote sensing and deep learning

This study utilizes drone-based remote sensing and deep learning to reveal that sympatric stingrays in a Red Sea lagoon exhibit fine-scale habitat partitioning and distinct foraging strategies, with bluespotted ribbontail rays occupying shallow mangrove edges for disruptive digging while larger whiprays inhabit macroalgal reef flats for non-disruptive feeding.

The Gist

Imagine the ocean floor near the coast as a bustling, underwater city. In this city, different groups of residents (fish and rays) need to find food and avoid predators without stepping on each other's toes. This is a concept ecologists call "habitat partitioning"—basically, everyone agreeing to live in different neighborhoods so they don't all fight over the same resources.

This paper is like a high-tech detective story about how two specific groups of stingrays—the tiny, colorful Bluespotted Ribbontail Rays and the larger, more serious Whiprays—manage to coexist in the same Red Sea lagoon without causing chaos.

Here is the breakdown of their story, told simply:

1. The Problem: Too Many Eyes, Not Enough Time

For a long time, scientists trying to watch these rays had to use expensive underwater cameras or tag the animals with trackers. It was like trying to watch a whole football game by only looking at one player at a time. You'd miss the big picture, and you could only watch a few players before your battery died.

Then, drones arrived. Think of a drone as a super-powered, silent bird hovering 35 meters (about 10 stories) up. It can snap thousands of high-definition photos of the shallow water in minutes. But here's the catch: The drone took so many photos that it created a "data mountain." If a human tried to look through every single photo to count the rays, it would take years. It's like being asked to find a specific grain of sand in a bucket of sand by hand.

2. The Solution: The AI "Super-Scanner"

To solve the mountain of photos, the researchers taught a computer an Artificial Intelligence (AI) trick. They used a smart system (called YOLO-11) that acts like a super-fast, tireless security guard.

• The Human vs. The Robot: The researchers tested this by having humans look at the photos and then letting the AI look at the same photos.
  • Humans found about 76% of the rays. They got tired, their eyes glazed over, and they missed the tiny ones hiding in the sand.
  • The AI found 97% of the rays. It didn't get tired, and it could spot a tiny, camouflaged ray that a human would walk right past.
• The Analogy: Imagine looking for a needle in a haystack. A human might miss a few needles because they are small and blend in. The AI is like a metal detector that beeps every time it finds a needle, no matter how small or hidden.

3. The Discovery: Two Different Neighborhoods

Once the AI did the heavy lifting, the researchers could finally see the pattern. They discovered that these two types of rays are basically "roommates" who have agreed to live in completely different parts of the house.

• The Bluespotted Rays (The Tiny Toddlers):

  • Where they live: They stick to the ultra-shallow water right next to the mangrove trees (less than 1.5 feet deep). It's like the "shallow end of the pool."
  • What they do: They are messy eaters. They dig into the sand with their mouths, creating little dust clouds to find worms and bugs hiding underneath.
  • Why? They are small and vulnerable. The shallow water acts as a shield against big sharks, and the mangroves are a buffet of easy-to-dig-up snacks.
  • The Impact: Because they dig so much in one specific spot, they turn that area into a "bioturbation hotspot"—basically, they are constantly remodeling the neighborhood, which changes how nutrients flow in the water.

• The Whiprays (The Big Teens):

  • Where they live: They prefer the slightly deeper water (about 2 feet deep) on the open reef flats, away from the mangroves.
  • What they do: They are polite eaters. Instead of digging, they glide over the top of the sand or seaweed, picking up food that is sitting on the surface (like crabs or shrimp). They don't make a mess.
  • Why? They are bigger and tougher. They don't need the shallow water as a hiding spot from predators, so they can roam further out.

4. The Big Picture: Why This Matters

This study shows that even though these rays live in the same small lagoon, they have figured out a perfect system to avoid fighting:

1. Different Addresses: One group stays in the shallow mangroves; the other stays on the open reef.
2. Different Jobs: One group digs up the floor; the other picks up what's on top.

This separation allows them to live together peacefully. If they didn't have these different habits, they would compete for the same food and space, and one group might die out.

The Takeaway

This paper is a win for science in two ways:

1. Ecology: It proves that we can see very small, detailed patterns in nature (like who eats what and where) that we couldn't see before. It shows us how nature organizes itself into tiny, efficient neighborhoods.
2. Technology: It proves that combining drones (to see the big picture) with AI (to count the details) is the future of conservation. It allows us to protect these animals better because we finally understand exactly where they live and what they need.

In short, the researchers used a flying camera and a robot brain to solve a mystery that human eyes alone couldn't crack, revealing that even in a crowded underwater city, everyone has their own special place to call home.

Technical Summary

1. Problem Statement

• Ecological Context: Sympatric species (species living in the same geographic area) often coexist through niche partitioning (differences in time, space, or diet). However, fine-scale habitat partitioning among mesopredatory elasmobranchs (rays and sharks) in coastal ecosystems remains poorly understood due to methodological limitations.
• Methodological Bottleneck: Traditional monitoring methods (acoustic/satellite telemetry) are expensive, have low spatial resolution, and struggle in very shallow waters. Conversely, Unoccupied Aerial Systems (drones) offer high-resolution data but generate massive volumes of imagery that overwhelm manual annotation capabilities.
• Specific Challenge: Detecting small, cryptic marine animals (like the bluespotted ribbontail ray, Taeniura lymma) in optically complex environments (waves, sun glint, caustics) is difficult for both human analysts and standard computer vision models, which often fail to detect small objects when downscaling images.

2. Methodology

The study employed a multi-faceted approach combining field surveys, active tracking, and advanced AI workflows in a coastal lagoon in the central Red Sea.

A. Data Collection

• Study Area: A ~17 ha coastal lagoon and adjacent reef flat (22.34° N, 39.09° E), featuring diverse benthic habitats (sand, macroalgae, mangroves, reef).
• Transect Surveys: 30 aerial surveys were conducted using a DJI Matrice 300 RTK drone with a 45MP Zenmuse P1 camera.
  • Flight Parameters: 35m altitude, 6 m/s speed, 70% frontal overlap, Ground Sampling Distance (GSD) of 0.41 cm/px.
  • Output: ~1,074 images per survey covering ~17 ha.
• Active Tracking: Manual drone tracking of 41 individuals (40 rays, 1 shark) for a total of >23 hours. A relay system using two drones extended tracking beyond battery limits.

B. Image Analysis & AI Workflow

• Double-Observer Protocol: Images were processed via two parallel streams: manual annotation and AI-assisted detection.
• Model Architecture: A custom-trained YOLOv11x model.
• Key Technical Innovations:
  1. Training Data: The model was trained on high-resolution imagery from the active tracking surveys to ensure visual consistency with the transect data.
  2. Slicing Aided Hyper Inference (SAHI): To address the "small object" problem (rays ≤0.35m in a 35m swath), the authors implemented a custom workflow. Original images (8192 × 5460 px) were sliced into overlapping tiles (2432 × 1632 px) for inference, and training images were cropped to 1344 × 1344 px. This allowed the model to operate at native resolution without losing detail.
  3. Post-Processing: Model detections (confidence threshold 0.5) were manually reviewed to remove false positives and duplicates caused by tile overlap.
• Statistical Modeling: Generalized Additive Mixed Models (GAMMs) were used to analyze presence/absence data against environmental gradients (depth, habitat type, distance to mangroves/open water, and conspecifics).

3. Key Results

A. Detection Performance

• Volume: 1,468 ray detections (6 species) and 4 sharks were recorded across transects.
• AI vs. Human: The AI-assisted workflow detected 97% of all observations, whereas human analysts detected only 76%.
• Efficiency: Manual annotation took up to 15 hours per survey; AI-assisted review took <1 hour.
• Accuracy: The model achieved a mean Average Precision (mAP-0.5) of 0.85. It successfully identified cryptic small rays that humans missed due to wave distortion and low contrast.

B. Ecological Findings: Fine-Scale Habitat Partitioning
The study revealed distinct spatial segregation between the two dominant taxa at sub-kilometer scales:

1. Bluespotted Ribbontail Rays (Taeniura lymma):
   • Habitat: Strictly occupied ultra-shallow lagoonal sandflats (<0.4 m depth), specifically adjacent to mangroves.
   • Behavior: High frequency of excavation feeding (digging), creating sediment plumes. Digging was concentrated near mangroves, identifying this zone as a bioturbation hotspot.
   • Drivers: Likely driven by prey availability (infaunal invertebrates) and predator avoidance (using shallow water as a refuge).
2. Whiprays (primarily Himantura uarnak):
   • Habitat: Predominantly occupied the adjacent reef flat with macroalgal habitats (~0.6 m depth).
   • Behavior: Foraged non-disruptively (epifaunal feeding) with minimal sediment disturbance.
   • Drivers: Less constrained by shallow-water stranding risks, allowing use of deeper reef flats.

C. Behavioral Observations

• Aggregation: Both taxa showed significant aggregation with conspecifics.
• Symbiosis: Needlefish were frequently observed following whiprays, and small fish fed on resuspended material around digging bluespotted rays.
• Site Fidelity: Unique individuals (identified by spot patterns or deformities) were re-sighted across months, demonstrating high site fidelity.

4. Key Contributions

• Methodological Advancement: Demonstrated a robust pipeline for detecting small marine animals in drone imagery using SAHI (Slicing Aided Hyper Inference) with YOLOv11. This overcomes the common limitation of downscaling images, which obscures small targets.
• Efficiency Scale: Proved that AI can reduce annotation time by >90% while simultaneously increasing detection rates by ~25% compared to manual methods.
• Ecological Insight: Provided empirical evidence that sympatric mesopredators partition habitat at very fine scales (<1 km) based on foraging strategy (excavation vs. surface feeding) and body-size constraints (stranding risk vs. predation risk).
• Data Availability: The authors released a dataset of ~110,000 images (with 3,681 annotated) to the public, addressing the scarcity of high-quality annotated marine datasets.

5. Significance

• Conservation & Management: The study highlights how the loss of top predators (sharks) may lead to mesopredator release, altering ecosystem functioning through concentrated bioturbation by rays. Understanding these fine-scale dynamics is crucial for Marine Protected Area (MPA) design.
• Ecosystem Engineering: The research identifies specific "bioturbation hotspots" where ray foraging significantly alters sediment structure and nutrient fluxes, acting as a key ecological process in coastal lagoons.
• Future Monitoring: The integration of drones and AI offers a scalable, non-invasive solution for long-term population monitoring and ecosystem-based management in shallow tropical waters, a region with high biodiversity and conservation risk.
• Limitations: The study notes that species-level identification of whiprays was limited by water clarity, and diel (day/night) patterns were not captured as flights were restricted to morning hours to minimize glare.

In conclusion, this paper establishes that combining high-resolution drone remote sensing with specialized deep learning workflows can reveal previously invisible ecological patterns, fundamentally changing how we monitor and understand the spatial ecology of coastal elasmobranchs.