Varietal classification and morphological traits prediction of lettuce cultivar (Lactuca sativa L.) based on RGB-images and convolutional neural network

This study evaluated five convolutional neural network (CNN) architectures for non-destructive lettuce (Lactuca sativa L.) phenotyping using RGB images. A dataset of 388 plants representing four cultivars was used to predict leaf area (LA), canopy diameter (D), fresh weight (FW), and dry weight (DW), and to classify cultivar identity. Images were split into training, validation, and test sets (70/15/15%) before training-only augmentation. An input-representation ablation compared original RGB, background-segmented RGB, and binary canopy masks. The best regression results were obtained with InceptionResNetV2 using original RGB for FW (R 2 = 0.935, RMSE = 30.65 g) and D (R 2 = 0.892, RMSE = 2.17 cm), SimpleCNN using original RGB for DW (R 2 = 0.958, RMSE = 1.02 g), and InceptionResNetV2 using segmented RGB for LA (R 2 = 0.898, RMSE = 520.73 cm 2 ). Cultivar-classification accuracy ranged from 96.55% to 98.28% across architectures. Five-fold cultivar-stratified cross-validation using segmented RGB showed relatively stable regression performance across partitions. Under the same data split, the best CNN configurations produced numerically higher R 2 values than feature-engineered machine-learning baselines, although the difference for LA was marginal. These results demonstrate trait-dependent effects of image representation and support RGB-based CNNs as a practical approach for lettuce phenotyping.