Abstract
Deep-learning pipelines that detect small ground objects (10–30 px) in single uncrewed aerial system (UAS) frames usually report positional accuracy as one end-to-end number against field references. We show that, at standard survey overlap, every object detected in several frames yields repeated position measurements, turning an operational campaign into its own positional test field: the precision of direct georeferencing can be measured, decomposed, and monitored without ground truth. Applied to 12,745 raw detections from 22 non-RTK DJI Matrice 30T surveys (European hamster burrow monitoring), 1307 multiply detected objects at 75 m above ground yield a median single-view deviation from the object centroid of 1.20 m (Q95 3.2 m); because views of one object often share a flight line, a time-stratified pairwise analysis places the full single-view error across lines at ≈1.6 m. The dispersion recovers known error structure: metre-class GNSS noise decorrelating within ~15 s; a within-line along-track component (σ 1.1 m along vs. 0.2 m across) consistent with the ≈0.3 s camera–GNSS time offset previously measured on DJI platforms; and a small radial image effect (0.018 m per metre of ground radius) whose signature points to height above ground rather than yaw error. The practical yield is a data-driven rule: set validation and matching tolerances at the Q95 of the measured dispersion (≈3.5–4 m). Anonymised data and code are available on request.