Model Limitations and Observations
Testing showed that detection performance was sensitive to lighting variation, surface glare, and background clutter. Thin debris and water reflections were the most common sources of false positives, particularly in low-contrast conditions. These results highlighted challenges in using raw image data for aquatic monitoring and motivated recommendations for improved lighting control, preprocessing, and expanded training data in future iterations.
Engineering Tradeoffs and Design Context
The system was designed around low-cost, readily available components to support affordability and ease of deployment. Design decisions favored portability, durability, and simplicity over image quality optimization, reflecting realistic constraints for field-deployable environmental monitoring devices. The project demonstrates the tradeoffs involved in integrating machine learning with embedded hardware in uncontrolled environments.