MulMatch-CL: A Semi-Supervised Teacher-Student Framework for Robust Crop Segmentation in UAV Imagery

Abstract

Accurate crop segmentation using unmanned aerial vehicle (UAV) imagery is essential for efficient crop monitoring and management. While Transformer-based architectures have demonstrated exceptional performance in segmentation tasks, their application to UAV imagery remains challenging owing to limited labeled data and noisy annotations. To address these challenges, this study proposes MulMatch-CL, a novel teacher-student architecture that integrates the semi-supervised MulMatch framework with Confident Learning (CL). The proposed MulMatch component employs consistency regularization and multiple strong augmentation streams to effectively utilize unlabeled data and enhance model generalization. A teacher model is first trained on both labeled and unlabeled data to generate pixel-wise probability distributions. Confident Learning then identifies and filters noisy labels, refining the labeled dataset. The cleaned dataset, combined with the unlabeled data, is used to train a student model, resulting in a more robust segmentation performance. Experimental results on the Barley Remote Sensing Dataset show that MulMatch-CL achieves 78.42% mIoU, 89.10% pixel Acc, and 87.58% F1 score, outperforming supervised baselines, robust learning strategies, and semi-supervised methods. Ablation studies further confirm that both Confident Learning and MulMatch independently enhance performance, improving mIoU by 2.36% and 4.28% respectively, while their integration yields a 6.08% improvement over the baseline. These results demonstrate that MulMatch-CL provides a robust solution for applying Transformer models to UAV-based crop segmentation.