Dimas Adrian Arifin, Nola Ritha, Nurul Hayaty · Springer Link (Chiba Institute of Technology) 2026 · 2026
DOI: 10.1051/bioconf/202626203004
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Mangrove species identification from field images is challenged by morphological similarity, occlusion, and complex backgrounds. This study evaluates EfficientDet-D0, RF-DETR-Small, and their weighted boxes fusion (WBF) ensemble on 8,100 images representing nine species and three organs (leaf, flower, and root; 27 classes). Under an image-level 70:20:10 split, RF-DETR-Small outperformed EfficientDet-D0 on the three reported aggregate mAP metrics. WBF achieved the highest mAP@0.5 (0.874 versus 0.869 for RF-DETR-Small), but RF-DETR-Small retained higher mAP@0.75 and mAP@0.5:0.95. The largest organ-level WBF gain over RF-DETR-Small occurred for roots at IoU=0.5 (0.678 versus 0.667). In 2,700 Monte Carlo trials constructed by pairing test images of the same known species, three-organ aggregation achieved 99.9–100.0% identification accuracy, compared with 99.7% for flowers alone and 88.1–93.1% for roots alone. These simulated results do not establish identification accuracy on independent physical plants. Overall, WBF provided a small AP@0.5 improvement while reducing performance under stricter localization criteria. Plant-disjoint evaluation and additional organ-combination experiments are needed to establish field generalization and the incremental value of each organ.
No comments yet — start the discussion below.