Ensemble YOLO architectures and UAV imagery for robust on-tree durian detection

Information from the abstract

Reliable on-tree detection of durians is a prerequisite for automated fruit counting, yield estimation and harvest planning, which still rely on time-consuming manual inspection. This study benchmarks lightweight deep learning detectors for on-tree durian detection in high-resolution imagery from unmanned aerial vehicles (UAVs). Five nano- and tiny-scale You Only Look Once (YOLO) variants (YOLOv8n, YOLOv9t, YOLOv10n, YOLOv11n and YOLOv12n) were evaluated under four cumulative configurations on 1,344 annotated images (11,205 fruit instances) from commercial orchards in Southern Thailand and public sources. Data augmentation and image enhancement each changed mAP@0.5 by at most 0.6 percentage points per model; because each configuration was trained once with fixed random seeds, these increments may lie within seed-to-seed variation and are reported as indicative only. YOLOv9t was the most accurate single model (94.0% mAP@0.5; 78.3% mAP@0.5–0.95). A weighted-box-fusion ensemble of YOLOv9t, YOLOv8n and YOLOv11n with test-time augmentation (TTA) reached 96.0% mAP@0.5, but its mAP@0.5–0.95 fell to 74.6% and its end-to-end latency rose to about 60 ms per image ( ∼ 17 frames per second, FPS) with parallel execution, compared with 6.5–17.4 ms per forward pass ( ∼ 57–154 FPS, excluding enhancement) for single models on a workstation GPU. Single models are therefore the natural choice for onboard or edge processing, and the ensemble for offline analysis. The detector was also coupled with the BoT-SORT tracker in a prototype video-counting workflow; this counting stage is demonstrated only qualitatively and has not yet been validated against video-level ground truth. The annotated dataset is publicly released.

Why this record is monitored

This record has an Impact Signal of 75/100 based on recency, source, collaboration, and bibliographic signals. It prioritizes monitoring and is not a judgment of research quality.

Related topics: Smart Agriculture and AI · Advanced Neural Network Applications · Industrial Vision Systems and Defect Detection