基于U-Net++的部分可见标签辅助监督作物–杂草分割

U-Net++-Based Crop-Weed Segmentation with Partial-Visibility Label-Assisted Supervision

摘要: 作物–杂草像素级分割是实现变量除草和田间精细管理的重要基础。针对PhenoBench数据集中部分可见作物(partial-crop)和部分可见杂草(partial-weed)在常规三分类合并后可见性信息被弱化的问题，本文以U-Net++为统一骨干，比较三分类直接监督(B0)、五分类直接监督(B1)、部分可见像元渐进式重加权(B2)和训练期五分类辅助监督(B3) 4种标签利用策略。四种方法均采用3个随机种子在官方训练集上训练，并在完整官方验证集上评价。结果表明，B3的平均交并比(mean intersection over union, mIoU)、杂草召回率和部分可见杂草召回率分别为0.7943 ± 0.0131、0.6170 ± 0.0117和0.5413 ± 0.0045。相较B0，B3的杂草召回率提高2.30个百分点，但杂草精确率下降2.88个百分点，mIoU仅提高0.24个百分点。B1在普通五分类交叉熵(cross-entropy, CE)与Dice联合损失下，部分可见杂草召回率较低且种子间波动较大。本文结果仅反映验证集上的描述性差异，总体而言，训练期五分类辅助监督能够在不改变三分类推理输出的前提下改善部分可见杂草的检出，但其精确率–召回率权衡及跨场景泛化仍需进一步验证。

Abstract: Pixel-level crop-weed segmentation is fundamental to variable-rate weeding and fine-scale field management. To address the weakened visibility information caused by merging partially visible crop (partial-crop) and partially visible weed (partial-weed) labels into a conventional three-class task in the PhenoBench dataset, this study uses U-Net++ as a unified backbone and compares four label-utilization strategies: direct three-class supervision (B0), direct five-class supervision (B1), progressive reweighting of partially visible pixels (B2), and training-only five-class auxiliary supervision (B3). All four methods are trained on the official training set using three random seeds and evaluated on the complete official validation set. The results show that B3 achieves a mean intersection over union (mIoU), weed recall, and partial-weed recall of 0.7943 ± 0.0131, 0.6170 ± 0.0117, and 0.5413 ± 0.0045, respectively. Compared with B0, B3 increases weed recall by 2.30 percentage points, while weed precision decreases by 2.88 percentage points and mIoU increases by only 0.24 percentage points. Under the five-class cross-entropy (CE) and Dice joint-loss configuration without class-balancing strategies, B1 exhibits lower partial-weed recall and greater variation across random seeds. The reported results represent descriptive differences on the validation set only. Overall, training-only five-class auxiliary supervision can improve the detection of partially visible weeds without changing the three-class inference output, although its precision-recall trade-off and cross-scenario generalization require further validation.