The precise segmentation of agricultural remote sensing images is pivotal for the effective monitoring and management of cultivated land resources. Traditional approaches often fall short in accurately delineating agricultural areas, hampered by complex surface features and the presence of non-arable elements such as buildings, roads, and wastelands. Addressing these issues, our study introduces a novel method for the unsupervised segmentation of agricultural remote sensing images through iterative application of the Segment Anything Model(SAM). Additionally, we significantly enhanced a classification model improved from ResNet50 by integrating two attention mechanisms. By integrating the SAM with classification model, we innovatively tackle the fundamental challenge of SAM-its segmentation results lack labels-thereby providing an inspiring solution for the entire field of image segmentation, including remote sensing imagery. This enhancement also enables precise instance segmentation for specific category, effectively retaining instances of cultivated land, which are more valuable for further agricultural analysis. Experiments demonstrate the high-efficiency of our proposed LiSegAgr, which has achieved state-of-the-art(SOTA) performance in the domain of agricultural remote sensing image segmentation. Compared to the classic instance segmentation algorithm Mask R-CNN, our method enhances the IoU metric by 17.2%. The code of our method is available on GitHub: https://github.com/WangYunKa/ISAgrSC2 .