Answers are grounded in this paper and cite their sources.
Estimating wheat head density from unmanned aerial vehicle (UAV) imagery supports field phenotyping; however, detection still faces the challenges of dense canopies and small targets. We propose AgriMamba-YOLO, a CNN–SSM detector that retains shallow convolutional features and applies four-direction selective scanning within each deep channel group. Target-specific P2–P4 fusion incorporates stride-32 context, normalized Wasserstein distance regression, and Gaussian Soft-NMS. Replacement and component controls evaluate the hybrid configuration, scale weighting, prediction resolution, and regression settings. The model achieves 64.6% AP50:95 on GWHD 2021, 4.5 points above YOLO11n. Five UAV runs give mean AP50 of 86.3% and AP50:95 of 61.2% on 364 native test frames; on the same GT, YOLO11n gives 53.9% AP50:95 and 86.5% AP50, a 7.25-point AP50:95 advantage for the proposed detector. Across 180 independently measured 1 m2 field quadrats, MAE is 12.4 and RMSE is 17.1 heads m−2. Complete-frame Jetson Orin Nano processing averages 1.66 s, with 0.8% of observations exceeding the 2 s capture interval. The evaluation connects held-out plot, later-date testing to field measurements, and device processing. Height group MAEs range from 10.5 to 14.0 heads m−2, while distinct quadrats across heights and two test plots on one date bound altitude and wider generalization inferences.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: