Growth-State Consistency Learning for Temporal Plant Growth-Stage and Leaf-Count Estimation

Purpose: This study addresses temporally inconsistent plant growth-stage predictions and visible leaf-count estimates from image sequences. Design/Methodology/Approach: Growth-State Consistency Learning (GSCL) combines an ImageNet-pretrained ConvNeXt-Tiny encoder, spatial self-attention, a two-layer temporal Transformer, growth-stage classification, visible leaf-count regression, and one-sided prediction-level monotonicity regularizers. PhenoLeaf-TS was partitioned by plant into 222 training, 48 validation, and 48 held-out test plants. Four-frame windows were used. Research Limitation: Since the stage labels are defined by visible leaf count thresholds, the two prediction tasks are statistically and conceptually coupled. In the setting where true leaf loss or senescence occurs, the prior may not be valid. Findings: The average predictions for 2514 unique test images resulted in an accuracy of 93.20% with 92.70% balanced accuracy and 92.90% macro-F1 score. The mean absolute error for leaf count was 0.69 leaves. The complementary window-conditioned evaluation showed 93.00% accuracy with 92.40% balanced accuracy and 92.60% macro-F1 score and a mean absolute error of 0.70 leaves. Practical Implication: The approach makes repeated-image monitoring more stable. Social Implication: More stable plant image monitoring could promote data-informed crop and plant research. Originality/Value: GSCL integrates temporal multi-task prediction with soft consistency regularisation on predicted outputs, reporting both unique-frame and overlapping-window evaluations.