Bench Characterization of Lightweight Object-Detection Models on an Edge-AI Camera for UAV-Oriented Source-Water Monitoring

A post-flight analysis of unmanned aerial vehicle (UAV) imagery has the potential to result in a delay in the inspection of source water. This delay can occur when visible debris or changes in the water surface necessitate a prompt response. The present study does not evaluate in-flight operation; rather, it presents a bench-level feasibility assessment of two deployment tasks—broad two-class screening and close-range debris classification—using lightweight YOLO detectors on an edge-AI camera in a host-fed configuration that approximates the timing constraints of a future UAV workflow. The YOLOv8, YOLO11, and YOLO26 models were lightweighted through structural pruning (YOLOv8) or architecture scaling (YOLO11 and YOLO26). These models were then refined through a process of fine-tuning, exported to the camera, and evaluated in terms of several metrics. The metrics encompassed training-environment accuracy, the accuracy of device-returned outputs, round-trip latency, and snapshot-based operating-load estimates. The dataset under consideration is extensive, comprising 4813 training images and 575 validation images, accompanied by 13,051 and 1615 annotations, respectively. The depth-pruned YOLOv8s variant demonstrated a significant reduction in mean round-trip latency, from 426.87 milliseconds to 231.58 milliseconds (45.75%), while the mAP@0.5 metric exhibited a decrease from 0.7018 to 0.6650, and the mAP@0.5:0.95 metric demonstrated a decline from 0.5433 to 0.5290. A class-level analysis reveals that aggregate accuracy is primarily influenced by the weaker floating-debris class, whose AP@0.5 ranges from 0.29 to 0.46, in contrast to the 0.82 to 0.94 range observed for pond/reservoir. In comparison to a matched baseline that was trained for an equivalent number of epochs with the sampler disabled, debris-biased sampling contributes 1.5 ± 0.6 mAP@0.5 points for YOLO11 and 3.6 ± 0.2 points for YOLO26 across three seed-matched pairs. The primary effect of this method is to increase floating-debris recall by 4.7–5.9 percentage points, with a concomitant small reduction in precision. The latency reduction increased the broad-inspection rate by 1.85×, provided approximately 195 milliseconds of idle margin within a 1-hertz cycle, and increased the paired far/near rate by 1.59× with two models resident on the camera. Three-seed repetitions of compact-model fine-tuning yielded 0.6717 ± 0.0033 and 0.6290 ± 0.0028 mAP@0.5. These results express detector compression in terms of operational monitoring capacity rather than model-size reduction alone, while also showing that compression by itself does not resolve the weak-class limitation that governs source-water inspection accuracy.