# Dark Vessel Detection Using ML at Defense Innovation Unit ## Project Overview The Defense Innovation Unit (DIU) developed a machine learning system to detect illegal, unreported, and unregulated (IUU) fishing activities using satellite-based synthetic aperture radar (SAR) imagery. This project addresses a critical global challenge, as one in five fish caught worldwide come from illegal fishing operations. The system was developed in collaboration with multiple partners including Global Fishing Watch, Coast Guard, NOAA, and academic researchers. ## Data Collection and Processing Challenges - Built large-scale dataset of SAR imagery - Complex data characteristics - Data annotation challenges ## Technical MLOps Challenges ### Model Architecture & Training - Baseline model used Faster R-CNN architecture - Required handling extremely large image sizes (20,000x20,000 pixels) - Long-range context preservation was critical ### Deployment Challenges - Efficient inference requirements - Edge deployment considerations ### Monitoring and Evaluation - Complex evaluation metrics - Operational monitoring ## Production System Features ### Data Pipeline - Automated ingestion of SAR imagery from multiple satellite constellations - Pre-processing pipeline for radar data normalization - Co-registration of multiple data sources (bathymetry, wind data, etc.) ### Model Deployment Architecture - Integration with SeaVision platform - Serving system handling large-scale inference - Queuing system to optimize cost and processing efficiency - Distribution to hundreds of countries worldwide ### Performance Results - Successfully deployed across 100+ countries - Identifies dark vessels not broadcasting AIS signals - Enables targeted enforcement in marine protected areas - Demonstrates ability to detect small vessels human analysts might miss ## Lessons Learned & Best Practices - Importance of stakeholder engagement and requirements gathering - Value of starting with open-source/public data before scaling to sensitive systems - Need for careful consideration of compute/power constraints in edge deployments - Benefits of iterative development with clear metrics and evaluation criteria - Importance of human-in-the-loop workflow design for critical decisions ## Impact & Results - System deployed globally through SeaVision platform - Enables better enforcement of marine protected areas - Helps countries with limited resources monitor their exclusive economic zones - Creates deterrent effect for illegal fishing activities - Demonstrates successful public-private partnership in AI deployment