English | EN 
Monitoring marine species is critical for protecting biodiversity, understanding ocean ecosystems, and guiding conservation efforts.
Today, artificial intelligence (AI) is transforming this field by enabling the automatic detection and counting of marine animals from images and videos at a scale that was previously impossible.
But while the promise is huge, building reliable AI systems for marine monitoring is far from straightforward.
A Complex Reality Beneath the Surface
Marine data comes from a wide range of sources:
- Underwater cameras
- Aerial drones
- Satellite imagery
Each of these technologies captures the ocean in a completely different way and that’s where the challenge begins.
The Core Problem: AI Doesn’t Generalize Easily
One of the biggest limitations of AI systems is their difficulty in adapting to new environments.
For example:
- A model trained on underwater fish images may fail on drone footage
- A system trained on drone images may struggle with satellite data
This phenomenon is known as domain shift when the data used for training differs from real-world conditions.
In marine monitoring, domain shift is everywhere.
Why Marine Data Is So Complex
1. Different Ways of Seeing the Ocean
Each imaging source has its own unique characteristics:
- Underwater images → blurry, low contrast, color distortion
- Drone images → small objects, changing angles, moving water
- Satellite images → massive scale, tiny objects, limited detail
Even when observing the same animal, its appearance can vary dramatically depending on how it was captured.
For an AI model, this is like recognizing the same person in:
- a selfie
- a drone shot
- a satellite image
Not so easy.
2. Detecting Very Small Animals
Scale is a major challenge:
- Underwater → fish may occupy most of the frame
- Drone → fish appear very small
- Satellite → fish may be just a few pixels
Small objects are particularly difficult to detect because:
- They contain very little visual information
- They blend into the background
- They can be confused with noise (waves, reflections, shadows)
3. Lack of High-Quality Data
AI models require large amounts of labeled data but in marine environments:
- Data collection is expensive
- Underwater acquisition is technically difficult
- Annotation requires expert knowledge
And it gets worse:
- Rare or endangered species have very few examples
- Datasets are often highly imbalanced
This severely limits real-world performance.
4. Heterogeneous Data Sources
Combining datasets from different organizations introduces additional complexity:
- Different image resolutions
- Different sensors
- Inconsistent annotation standards
- Variable data quality
The result: heterogeneous datasets that are hard for models to learn from.
5. Limited Transferability of AI Models
Deep learning models tend to learn patterns that are very specific to their training data.
When conditions change such as:
- Lighting
- Viewpoint
- Environment
👉 The model may fail completely.
Even advanced techniques designed to align different data domains often struggle when differences are too large for example, between underwater imagery and satellite data.
6. Synthetic Data Isn’t a Silver Bullet
To compensate for data scarcity, researchers sometimes generate synthetic datasets.
However:
- Simulated underwater scenes don’t fully capture real-world complexity
- Synthetic aerial or satellite images may lack realism
This creates a gap between simulation and reality, limiting model performance.
What This Means for Real-World Marine Monitoring
All these challenges point to a key reality:
Building a single AI system that works across underwater, aerial, and satellite data is extremely difficult.
Yet solving this problem is essential for:
- Tracking marine biodiversity
- Monitoring endangered species
- Supporting environmental policy decisions
- Automating large-scale ocean observation
Where AI Is Heading
Researchers are actively exploring new approaches to overcome these limitations:
- Domain Adaptation → adjusting models to new environments using available data
- Domain Generalization → designing models that work in unseen conditions
- Few-shot learning → learning from very small datasets
- Multimodal AI → combining multiple data sources
- Hybrid pipelines → integrating detection, classification, and reasoning (e.g., with LLMs)
Key Takeaway
AI has enormous potential to transform marine species monitoring but deploying it effectively in the real world remains a major challenge.
The core issue is no longer just detecting animals.
It’s about building systems that work reliably across different environments, sensors, and conditions.
Solving this will be a crucial step toward scalable, global, and trustworthy ocean monitoring systems.