Vision-Based Strawberry Ripeness Grading for Harvest Decision Support in Smart Greenhouses

Strawberry ripeness estimation is essential for determining optimal harvest timing and supporting quality management in smart greenhouse environments. However, since strawberry ripening is a continuous process rather than a transition between discrete visual boundaries, standard RGB-based methods often struggle to distinguish subtle color transitions during intermediate ripeness levels, and misclassifications among adjacent ripeness stages can substantially affect harvest decisions. This makes it important to not only achieve high classification accuracy but also to minimize the ordinal deviation between predicted and actual ripeness stages. To address these limitations, this study presents a two-stage vision-based framework for five-class strawberry ripeness grading. In the first stage, a YOLO11m-based detector localizes strawberry regions within greenhouse images and generates regions of interest (ROIs). In the second stage, each extracted ROI is classified using a five-channel input representation that combines RGB, Excess Green (ExG), and Hue information to enhance color-sensitive feature extraction. In addition, a specialized backbone, termed CP2-A2C2F, is designed to preserve high-resolution P2 features and integrate them with deeper P4 semantic representations through a dual-scale fusion strategy. This structure enables the model to capture localized chromatic variations and contextual maturity information simultaneously. Experimental results show that the CP2-A2C2F model achieved the best overall classification accuracy of 87.7% and a weighted F1 score of 0.877, together with the lowest mean absolute error (MAE) of 0.162 among models equipped with an auxiliary regression head, indicating the smallest average ordinal deviation between predicted and ground-truth ripeness stages. The ablation study further confirmed that adding ExG and Hue channels moderately reduced ordinal prediction error compared with the RGB-only configuration, and that the proposed P2 + P4 fusion achieved the best accuracy, weighted F1 score, and MAE among the evaluated fusion configurations, indicating that direct fusion of high-resolution P2 features with deeper P4 semantic features provides a more effective balance than including the intermediate P3 level. These results indicate that the combination of explicit color representation and dual-scale P2 + P4 feature fusion can improve five-class strawberry ripeness grading and help reduce ordinal prediction errors under the tested greenhouse conditions.