SCEA-YOLO: A General-Purpose Maturity Grading Model of Multi-Crop Greenhouse Robots.

Abstract

Accurate classification of fruit maturity is essential for automated grading and robotic manipulation in modern greenhouse cultivation. Most existing methods rely on crop-specific models, severely restricting their scalability in multi-crop scenarios. To overcome this limitation, this study presents SCEA-YOLO, a unified and efficient instance segmentation framework built on YOLOv11s-seg, for simultaneous maturity classification of tomatoes and sweet peppers. To boost feature discrimination, reduce computational redundancy, and alleviate class imbalance, SCEA-YOLO integrates spatial-channel reconstruction convolution and an efficient multi-scale attention mechanism, while replacing the original detection head with the proposed EA-Head. The model is evaluated on a hybrid dataset captured under diverse greenhouse conditions, including varying illumination, fruit occlusion, and overlapping canopies. Its robustness to different viewing angles and camera distances is further validated via deployment on an automated grading robot. Compared with the baseline, SCEA-YOLO enhances classification precision and mAP50–95 by 5.3% and 2.3% for tomatoes, and 1.2% and 1.4% for sweet peppers, respectively. With only 33.2 GFLOPs, the model satisfies real-time inference demands. Benefiting from its lightweight structure and real-time performance, SCEA-YOLO can be readily deployed on embedded systems and robotic platforms. It offers a practical, unified, and scalable solution for intelligent fruit maturity evaluation in multi-crop greenhouse production.

Plant phenotyping relevance

トマトとピーマン果実の成熟度を画像から分類・評価するモデルを開発し、データセットおよびロボット上で性能検証しており、植物表現型取得手法が中心である。

Code and data availability

The paper's tomato/sweet-pepper maturity image dataset and trained SCEA-YOLO model are not deposited in a public repository: the Data Availability Statement explicitly restricts access to on-request sharing. A large (133.7 MB) MDPI supplementary zip exists at the official s1 URL, but the text does not state that it is,

No evidence-backed public reproduction asset is currently recorded.