Self-supervised representation learning for robust weed detection

Manually annotating agricultural datasets is typically expensive and inefficient and often lacks the environmental diversity required for effective deep-learning models. This work examines the influence of SSL on the effectiveness of CNNs for weed detection under several environmental conditions. SimSiam was employed to pretrain models on 110,986 unlabeled field images collected across multiple environmental domains. The experimental results demonstrate that SSL pretraining produces semantically meaningful, collapse-free representation spaces and achieves higher transferability to our downstream task compared to traditional transfer learning from public datasets. In addition, the use of SSL-pretrained models demonstrated robustness to environmental domain shifts and long-tail events.