The growing demand for accurate information in precision agriculture, driven by climate change and the need for resource optimization, has led to the development of advanced techniques for satellite image segmentation. However, traditional segmentation models face generalization challenges due to the scarcity of labeled data. This work proposes a lightweight approach based on the Segment Anything Model (SAM), enhanced through an iterative pseudo-labeling strategy that relies on segmentation consistency under salt-and-pepper noise perturbations. The methodology, grounded in semi-supervised learning, is validated through experiments on real satellite imagery. Results demonstrate that this self-consistency criterion enables effective use of unlabeled data, yielding substantial improvements in segmentation performance over standard fine-tuned SAM.