Estimation of diameter, mass and volume of Lanzhou lily bulbs based on YOLO instance segmentation

The Lanzhou lily is the only sweet lily species cultivated in China, contributing approximately 20% to Lanzhou's total agricultural output value. In recent years, lily bulbs' quality issues such as declining single-clove rates and bulb shrinkage have emerged, making rapid and accurate assessment of bulb viability crucial for mitigating degeneration. To address the subjectivity and inefficiency of manual screening, this study proposes a novel method. This method is based on the YOLOv8-seg instance segmentation model and is designed for high-precision prediction of lily bulb mass, diameter, and volume. A dataset was constructed by randomly selecting 500 lily bulbs with different morphologies, capturing their images using a mobile phone, and acquiring ground truth data (mass, diameter, volume) with an electronic balance, a digital vernier caliper, and a measuring cylinder, respectively. The YOLOv8-seg model was employed to achieve instance segmentation of the bulbs, followed by extraction of 22 morphological features via OpenCV. Combining the Variance Inflation Factor (VIF), Principal Component Analysis (PCA), and Mutual Information (MI), retaining three key predictors. Subsequently, six machine learning models were constructed to predict bulb mass, diameter, and volume. The experimental results indicate that the artificial neural network (ANN) model has the strongest generalization ability. The coefficient of determination between the predicted values and the measured values for the diameter, volume, and mass of lily bulbs are 0.8478, 0.9844, and 0.9981, respectively. The performance of the ANN model on the test set is as follows: the mean square error (MSE) for diameter prediction is 8.3175 and the mean absolute error (MAE) is 2.3019; the MSE for volume prediction is 1.0912 and the MAE is 0.7951; the MSE for mass prediction is 0.3135 and the MAE is 0.4318. This method enables the direct, rapid prediction of bulb diameter, volume, and mass from RGB images, providing agricultural practitioners with an efficient and reliable tool for bulb viability detection.