An end-to-end maturity prediction and hierarchical counting method for cherry tomatoes

In vertical farming, accurately and automatically estimating cherry tomato maturity remains an unsolved challenge with significant potential to enhance agricultural efficiency and crop quality. This study addresses the complexities of maturity estimation and occlusion detection in greenhouse environments, where variable lighting and frequent occlusions complicate visual assessments. This work presents a Lightweight Cherry Tomato Multi-Task Detection (CT-MTD) model, a deep learning-based multi-task framework designed to perform maturity estimation and occlusion detection. The model utilises a shared MobileVIT backbone for efficient feature extraction, with task-specific branches for regression-based maturity estimation and classification-based occlusion detection. Coordinate Attention (CA) is integrated before each branch to capture spatial and channel-wise dependencies. To support precise data labelling, a custom annotation tool allowing horticultural experts to annotate maturity stages and occlusion statuses was developed. The experiments demonstrate that the CT-MTD model achieves a RMSE of 0.048 for maturity estimation, and Accuracy and F₁ scores of 0.817 and 0.778 for occlusion detection, validating the effectiveness of the MobileVIT and CA combination for multi-task learning in precision agriculture. Finally, to validate the maturity estimation algorithm in inspection robots, a fruit counting method for tomatoes at different maturity levels was developed. The mean maturity prediction accuracy for tomato counting in two greenhouses was 86.3 % and 80.2 %, demonstrating the model's robustness across different greenhouse conditions. This work can help greenhouse operators anticipate daily output, facilitating market forecasting and efficient logistics for transportation.