"Adaptive Modular Detection with Dynamic Decision Mechanism for Robust Soil Attachment Recognition on Angelica Roots"

"Accurate and real-time detection of soil attachment on \\textit{Angelica dahurica} roots is essential for stabilising post-harvest handling quality and energy consumption. In this work, a modular YOLOv11 framework and an Adaptive Decision and Module Selection (ADMS) mechanism are built with two pluggable slots (Slot~A and Slot~B) that host functionally different modules, including the lightweight downsampling unit ADown, the Attention-Scale Fusion (ASF) module, and structural balancing units LADH and FDPN. Static experiments show that the baseline YOLOv11 maintains the highest overall accuracy (Precision = 0.994, mAP$_{50}$ = 0.9947), YOLOv11-ADown achieves the best recall and real-time performance (Recall = 0.9981, FPS = 157.6), and YOLOv11-ASF provides the highest precision (Precision = 0.9866) with more stable feature responses. Built on this modular foundation, ADMS dynamically switches between Baseline, Slot~A, and Slot~B according to frame-wise performance and scene conditions; in video tests on newly captured and randomly concatenated field sequences, it selects the theoretically optimal module in 62.2\\% of Heavy Soil frames while still keeping non-zero usage of alternative modules in Light Soil and Normal scenes. Under recall-, precision-, and speed-priority weighting strategies, Slot~A consistently reaches the highest composite score $S_m$, and across six normal and extreme scenarios ADMS keeps detection continuity above 0.97 with fallback frequency below 0.8\\%. Overall, the proposed modular YOLOv11 with ADMS achieves a practical balance between accuracy and real-time performance and provides robust, scene-aware structural adaptation for agricultural visual perception. "