.
We built a framework that uses NASA satellite data and AI models to predict where sharks forage. Alongside, we designed a smart shark tag concept capable of detecting feeding events in real time.
This interactive web application demonstrates real-time shark tracking visualization combined with oceanographic data. It is built with HTML, CSS, and JavaScript, and offers both 2D map and 3D globe views for exploring marine ecosystems.
1. Dual Visualization Modes
2. Interactive Shark Tracking
3. Sidebar Controls (toggle with menu button)
This mockup demonstrates how satellite oceanographic data could be integrated with biologging telemetry to study predator-prey relationships and migration patterns in marine environments.
Framework that Infers Shark Hotspots
Controlled environment data gathering for Edge AI model training
Build a labeled dataset of feeding events in controlled environments to train and validate the behavioral detection model
Edge AI-powered behavioral detection with satellite communication
Detect feeding events on-board and transmit real-time behavioral data to continuously update predictive models
V1 collects training data in controlled settings
Edge AI model is trained and optimized
V2 deploys trained model for real-time ocean monitoring
The challenge was to create a mathematical framework for identifying sharks and predicting their foraging habitats using NASA satellite data. It also involved designing a conceptual smart tag capable of detecting not only where sharks are, but also what they’re eating, and transmitting that information in real time to improve predictive models.
The goal was to uncover foraging hotspots and quantify how ocean conditions, phytoplankton communities, and predator movements are connected. As part of this, we explored new ways satellite measurements could help identify these crucial shark habitats.
F.I.S.H. Project is a research initiative combining AI, mathematics, and satellite data to predict shark foraging hotspots and support marine conservation.
It consists of the Poisson Point Process (PPP) model enhanced with two inference strategies: a Gaussian Process for interpretability and uncertainty estimation, and a Transformer-based neural network for scalability and real-time learning.
Early results revealed that biogeochemical data alone can’t accurately predict shark behavior, emphasizing the need for behavioral insights. To fill this gap, we designed a conceptual smart shark tag equipped with Edge AI that can detect feeding events directly on the animal and transmit key data via satellite.
A web application complements the framework, visualizing real-time shark trajectories and feeding hotspots across the globe. Together, these innovations demonstrate how AI, edge computing, and satellite observations can work hand in hand to advance marine research.
The project is being developed in collaboration with the Department of Data Science at the University of Trieste, with plans for a scientific publication.
The project aims to predict shark foraging habitats by combining satellite observations, advanced mathematical modeling, and real-time Edge AI tracking. This helps protect ocean ecosystems, inform conservation efforts, and support sustainable fishing policies.
Yes! The code and visual outputs are available on GitHub and in the project documentation.
A PPP is a mathematical model for random events scattered in space and time. It’s perfect for shark sightings recorded as individual points (day, latitude, longitude). We use an inhomogeneous PPP where shark density varies based on environmental conditions like temperature, salinity, and chlorophyll levels.
Existing tags provide location data, but we need behavioral data, specifically, when and where sharks feed. Our Edge AI-enabled tag detects feeding events on-board in real-time, which is critical for identifying true foraging hotspots and improving prediction accuracy.
The tag uses accelerometer data (ODBA – Overall Dynamic Body Acceleration) segmented into 2-second windows at 25 Hz. A lightweight convolutional neural network (CNN) trained via TinyML classifies shark behaviors like resting, swimming, and feeding based on movement patterns across three axes.
It uses a hierarchical sensing approach:
https://pace.oceansciences.org/data_images_more.htm?id=523
https://pace.oceansciences.org/data_images_more.htm?id=531
https://pace.oceansciences.org/data_images_more.htm?id=528
https://modis.gsfc.nasa.gov/data
Ivana is a content & SEO Specialist with 5+ years of experience
Copyright © 2025 Ivana Soldat. All Rights Reserved