Leaf or Leaves? A Data-Centric Approach Looking at Ground Truth Effect on Apple Disease Segmentation

In image-based plant disease detection, semantic segmentation plays a significant role in understanding fine-grained images and separating their meaningful parts. While recent advances in supervised segmentation models have shown promise in leaf and disease identification, most efforts have focused on optimising model architectures and parameters. However, enhancing segmentation results through data-centric methods is another perspective that emphasises improving data quality and quantity to boost model performance. Due to the high cost of data annotation, inconsistency in labelling, and the impact of ground truth quality on segmentation outcomes, there is a need to explore the effect of annotation on segmentation results. This study addresses the challenge of achieving consistency in labelling required for the segmentation of healthy and unhealthy plant tissues by defining a rule to guide the labelling process. Three sets of masks were generated and the effect of annotation on the segmentation results was evaluated using three models (UNet, PSPNet, and DeeplabV3+). Evaluation results highlight the importance of data quality and quantity and suggest that increasing data by including more annotated objects leads to better performance only if they are clear and qualified.