Table 3 Comparative overview of the GWHD and Bios-Wheat datasets used in the study.
From: Transfer learning models for wheat ear detection on multi-source dataset
Feature | GWHD Dataset | BioS-Wheat Dataset |
|---|---|---|
Number of images | 3442 | 5696 |
Number of annotated objects | 197,606 | 54,485 |
Classes | 1 (wheat head) | 1 (wheat head) |
Annotation type | Axis-aligned bounding boxes | Axis-aligned bounding boxes |
Annotation precision | High; semi-automatic labeling refined manually; minor errors from occlusion, lighting, and perspective | High; manual verification; very dense plots with frequent occlusions and overlaps |
Field conditions | Outdoor wheat plots; sowing densities: low, normal, high; row spacing 12.5–30.5 cm | Outdoor wheat plots; extremely high sowing density (500–550 seeds per m2); minimal row spacing 12.5 cm |
Growth stage | Post-flowering (fully emerged ears) | Watery, ripe, and medium milk stages (BBCH scale) |
Viewpoint/perspective | Nadir or near-nadir; phenotyping platforms or hand-held cameras | Nadir; smartphone positioned 1 m above canopy with 0.5 m \(\times\) 0.5 m ROI frame |
Lighting conditions | Natural daylight, mostly morning hours; moderate variation in brightness | Morning hours, cloud-free days; minimized variation |
Background/clutter | Soil, grass, weeds; moderate occlusion and visual complexity | Dense wheat plots; frequent occlusion, overlapping heads, irregular ear orientations |
Image resolution / GSD | \(1024\times 1024\) px; GSD \(\approx 0.1 - 0.6\) mm depending on platform | \(4032\times 3024\) px; GSD 0.3 mm per pixel |
Camera/sensor types | Various RGB cameras on phenotyping platforms | Smartphone with triple-camera system (ultra-wide 13 mm, wide 26 mm, telephoto 52 mm) |
