Utilising Machine Learning to Detect Debris in Lakes and Rivers

Flew a drone over local water bodies (including Cypress Creek) and ran two pretrained object detectors, Darknet and YOLOv3 (no fine-tuning), to identify debris, then manually checked detections against what was actually in each scene. YOLOv3 reached ~32% accuracy vs. ~7% for Darknet; common materials like glass, plastic, and cans were identified far more reliably than irregular debris. Also tested robustness against Gaussian noise and partial/cropped images.