Hail-producing convective thunderstorms are a major threat to agriculture and infrastructure causing large financial losses. Remote sensing techniques such as dual-polarimetric weather radar can provide hail observations over large areas, but do not necessary reflect the situation on the ground. Current ground-based observations—such as automatic hail sensors, hail pads, and crowd-sourced reports—provide valuable information but exhibit limitations for validating radar products in terms of area coverage. Drone-based hail photogrammetry coupled with machine-learning (ML) techniques has the potential to close this observational gap by sampling thousands of hailstones within the hail core across large areas of hundreds of square meters and provide a hail size distribution estimation. However, the reliability of this new technique has not yet been assessed. In this study, we conducted experiments on different grass surfaces using synthetic hail objects of known sizes and quantity to assess the uncertainty of the ML-based hail size distribution retrievals. The findings of the experiments are then compared with a real hail event surveyed using drone-based hail photogrammetry. Using drone-based hail photogrammetry coupled with ML, 98% of the synthetic hail objects and 81% of hailstones were correctly detected. Additionally, sizes of the detected objects were retrieved with a minor underestimation of around −0.75 mm across all sizes for both synthetic hail objects (10–78 mm) and hailstones (3–24 mm). Hence, the high accuracy coupled with a large sampling area provides an estimation of representative hail size distributions on the ground. These reliable ground observations are a valuable basis for applications such as validation of weather radar hail estimates.