Abstract: Accurate cropland extraction from high-resolution remote sensing images is often required for cropland protection, agricultural resource management, and national food security. However, cropland parcels in high-resolution imagery are characterized by strong internal texture heterogeneity, narrow field ridges, fragmented spatial patterns, and irregular boundaries, leading to particularly challenging cropland extraction. Conventional pixel and object detections, as well as conventional deep learning semantic segmentation models, frequently suffer from internal fragmentation, boundary discontinuity, and parcel adhesion under complex agricultural landscapes. Although the segment anything model(SAM) demonstrated remarkable generalization over diverse visual domains, it is lacking in direct application to cropland extraction due to the lack of explicit constraints on agricultural texture noise and parcel boundary structures. In this study, a SAM-based cropland-constrained segmentation network (SFCSNet) was proposed for high-resolution remote sensing images. A frozen Vision Transformer image encoder from the SAM was adopted as the backbone in the network to preserve strong global semantic representation. A hierarchical adaptive smooth constraint feature fusion (HASCF) module was designed to mitigate texture-induced noise for intra-parcel semantic consistency. Gaussian smoothing constraints were integrated with an efficient channel attention (ECA) mechanism to adaptively filter irrelevant high-frequency textures for meaningful cropland structures during multi-level feature fusion. In addition, a multi-granularity edge refinement constraint (MGER) module was taken as an auxiliary supervision branch to explicitly model cropland boundaries. Scharr-based edge responses and multi-granularity boundary supervision were incorporated to guide the network toward accurate delineation of irregular parcel contours, particularly for low boundary breakage and parcel adhesion. The overall architecture followed an encoder–decoder paradigm. Joint optimization was combined with semantic segmentation supervision and edge refinement supervision during training. Only the semantic branch was retained during inference to avoid additional computational overhead. Extensive experiments were conducted on two public high-resolution cropland datasets, including the Gaofen-2-based GID-5 dataset and the JiLin-1 cropland dataset. Quantitative results demonstrated that the proposed method outperformed the mainstream segmentation models selected in this study, including typical convolutional neural network models, transformer-based models, and SAM-derived general segmentation models, represented by DeepLabV3+, UNetFormer, the original SAM and its variant SAMUS. On the GID-5 dataset, the proposed method achieved an intersection over union (IoU) of 83.00% and an F₁ score of 90.71%. On the JiLin-1 dataset, the IoU reached 90.05% and the F₁ score reached 94.76%, indicating strong robustness under different regional and imaging conditions. Ablation experiments further confirmed that the HASCF module effectively reduced internal holes and fragmentation caused by texture interference, while the MGER module significantly improved boundary continuity for low parcel adhesion, especially in narrow, elongated, and densely distributed cropland regions. Visual comparisons revealed that there were the smoother cropland interiors, more continuous boundaries, and fewer misclassified regions than competing approaches in complex agricultural scenes. Overall, the proposed method effectively enhanced both semantic consistency and geometric integrity of cropland extraction. The findings can provide a reliable and practical solution for high-resolution cropland extraction. Technical support can offer large-scale applications of remote sensing in modern agriculture.