Answers are grounded in this paper and cite their sources.
Unbroken and broken soybean seeds may exhibit different morphology–mass relationships, complicating breakage-rate estimation from two-dimensional (2D) images. This study combined YOLO11n-seg instance segmentation with seed-type-specific mass prediction for Qihuang 34 seeds at a measured moisture content of 16.1% (wet basis) under static laboratory conditions. Predicted masks provided area, perimeter, circularity, and rectangularity. Models used 120 seeds per type and were evaluated over 30 repeated outer splits with internal five-fold tuning. Feature ablation selected ordinary least squares with perimeter, circularity, and rectangularity for unbroken seeds, and radial basis function support vector regression with area, perimeter, and circularity for broken seeds. In independent tests of 100 seeds per type using true seed types and predicted masks, the models achieved R2 = 0.860 and 0.809, RMSE = 0.0128 and 0.0109 g, and MAE = 0.0102 and 0.0086 g for unbroken and broken seeds, respectively. Under the fully automatic end-to-end condition using predicted seed types and predicted masks, broken-seed mass prediction yielded R2 = 0.646, RMSE = 0.0148 g, and MAE = 0.0096 g. For 50 independent mixed-seed groups (408 seeds in total; 7–18 seeds per group; measured breakage rates of 2–37%), breakage-rate prediction yielded R2 = 0.997, RMSE = 0.526 percentage points, and MAE = 0.407 percentage points. For 20 additional low-breakage-rate and near-boundary groups (352 seeds in total; 16–20 seeds per group; measured breakage rates of 2.29–6.24%), RMSE and MAE were 0.215 and 0.183 percentage points, respectively, and all absolute errors were below 0.5 percentage points. The method was feasible under these conditions; transferability and real-time field performance require validation.
No takes yet. Share an insight, caveat, or question.
Nan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: