Class merging degrades learned representations in potato-harvesting object detection despite operational equivalence

In artificial intelligence (AI)-operated mobile potato harvesting, separating field impurities (stones and soil clods) from potatoes determines product quality. Because both types trigger the same pneumatic removal action, merging them into one impurity class is a natural simplification; we found that it degrades detection quality in every random seed tested, relative to a 3-label formulation (potato vs. stone vs. soil clod).

We trained YOLOX-Small under two label structures, with identical images, boxes, augmentation and hyper-parameters: 3-label (potato, stone, soil clod) and 2-label (potato, impurity). The models were fitted on 11,869 images of ten cultivars from seven Hokkaido farms collected up to the 2024 harvest - the training and validation splits of a 13,189-image corpus divided 7:2:1. The test set was a separate 2,960-image corpus from the 2025 season sharing no image or acquisition session with it (110,420 potato and 15,234 impurity instances; 12.1% prevalence). Each structure used five seeds. Both were scored on a common two-class target, potato versus impurity: the stone and soil-clod predictions of the 3-label model were pooled into a single impurity class before any metric was computed. Detection was assessed with potato misclassification rate (PMR), impurity detection rate (IDR), impurity precision and average precision (AP) over a 17×17 threshold grid; the 512-dimensional backbone representations were compared on the same target.

The 3-label structure achieved higher impurity precision than the 2-label structure at the default operating point (83.20 ± 0.85% vs. 70.07 ± 1.59%; +13.1 percentage points (pp); 95% confidence interval (CI) [10.5, 15.7]) and higher threshold-free impurity AP at 50% box overlap (79.8 ± 1.3% vs. 65.6 ± 0.8%; +14.2 pp), in every seed across the operational threshold region. The 2-label representation showed a 39% smaller centroid distance, 37% lower silhouette separability and 30-41% lower per-channel Fisher ratios. The advantage persisted although the subclass task failed asymmetrically: soil clods collapsed into stone (recall 1.84 ± 0.13%) while stones were correctly subclassed (80.69 ± 1.12%).

Fine-grained supervision appears to shape the backbone representation independently of subclass accuracy: operational equivalence does not entail representational equivalence. The link between representation and performance is observational, and all results come from one detector architecture and region.