Instance-level phenotype-based growth stage classification of basil in multi-plant environments

Abstract

Climate change, shrinking arable land, urbanization, and labor shortages increasingly threaten stable crop production, attracting growing attention toward AI-based indoor farming technologies. Accurate growth stage classification is essential for nutrient management, harvest scheduling, and quality improvement; however, conventional studies rely on time-based criteria, which do not adequately capture physiological changes and lack reproducibility. This study proposes a phenotyping-based and physiologically grounded growth stage classification pipeline for basil. Among various morphological traits, the number of leaf pairs emerging from the shoot apex was identified as a robust indicator, as it can be consistently observed regardless of environmental variations or leaf overlap. This trait enables non-destructive, real-time monitoring using only low-cost fixed cameras. The research employed top-view images captured under various artificial lighting conditions across seven growth chambers. YOLO automatically detected multiple plants, followed by K-means clustering to align positions and generate an individual dataset of crop images-leaf pairs. A regression model was then trained to predict leaf pair counts, which were subsequently converted into growth stages. Experimental results demonstrated that the YOLO model achieved high detection accuracy with mAP@0.5 = 0.995, while the A convolutional neural network regression model reached MAE of 0.13 and R² of 0.96 for leaf pair prediction. Final growth stage classification accuracy exceeded 98%, maintaining consistent performance in cross-validation. In conclusion, the proposed pipeline enables automated and precise growth monitoring in multi-plant environments such as plant factories. By relying on low-cost equipment, the pipeline provides a technological foundation for precision environmental control, labor reduction, and sustainable smart agriculture.

Keywords: BBCH scale; automated decision pipeline; controlled environment agriculture (CEA); smart agriculture; vision-based phenotyping.

Figure 1

Overview of the top-view image collection system in growth chambers.

Figure 2

Representative examples of basil (Ocimum basilicum L.) growth stages based on the number of visible leaf pairs from top-view images: (A) Level 1 (≤ 2 pairs), (B) Level 2 (3–4 pairs), and (C) Level 3 (≥ 5 pairs). Arrows indicate the apical bud with emerging leaf primordia used as reference points for stage classification.

Figure 3

Examples of excluded images during preprocessing. (A) Multiple plants within a single pot; (B) Images captured during the dark period; (C) Plants more than half out of frame; (D) Non-germinated pots; (E) Plants with leaf removal history.

Figure 4

Automated labeling pipeline for basil plants: (A) YOLOv8 detects individual plants, (B) the detected plants are grouped into rows using K-means clustering, (C) plants within each row are sorted from left to right, and (D) sequential numbering is assigned for consistent labeling.

Figure 5

Automated pipeline for generating cropped plant images with matched leaf count labels for CNN model training, using YOLOv8 detection and preprocessing.

Figure 6

Integrated inference pipeline for plant growth stage classification. YOLOv8 detects individual plants, ResNet-18 regression estimates leaf-pair counts, and predictions are converted into discrete growth stages.

Figure 7

(A) Confusion matrix for the validation set, showing 352 true positives, 0 false positives, and 1 false negative. (B) F1–confidence curve illustrating the optimal confidence threshold that maximizes the F1-score. (C) Training and validation curves for box loss, classification loss, distribution focal loss, precision, recall, and mAP@0.5/0.5–0.95 over 100 epochs (x-axis: epoch, y-axis: corresponding loss or metric value).

Figure 8

Model training and classification performance. (A) Training and validation loss curves showing stable convergence, with early stopping triggered at the 27th epoch to prevent overfitting; (B) Normalized confusion matrix for Level 1–3 classification on the test set, indicating high prediction accuracy across all levels.

Figure 9

Automated growth stage classification of basil plants. (A) Cultivation bed 1, showing detection and classification of individual plants based on leaf-pair counts; (B) Cultivation bed 4, demonstrating stable multi-plant detection and classification under artificial lighting. Bounding boxes denote individual plants, and labels indicate the predicted growth stage (Level 1–3) and leaf-pair number.