Cultivated land not only is the foundation and lifeline of food production, but also an important premise and basis for ensuring national food security. Accurate identification of farmland parcel is of great significance in agricultural production. This study proposed a dual-stream-fusion network (DSFNet), using a semantic segmentation network and a boundary detection network in parallel, for farmland parcel in high-resolution satellite images. We proposed a semantic segmentation module based on DeepLabV3+ network structure and a boundary detection module based on DexiNed network to extract richer farmland parcel features by fusing the semantic and boundary information. Then, a post-processing optimization scheme is designed based on morphology to improve the accuracy of farmland parcel. Moreover, the slicing and stitching of large remote sensing images are achieved by using the expansion prediction strategy to obtain the final high-precision result without obvious stitching trace. The experimental results illustrate the effectiveness of the proposed model for fine-scale farmland parcel mapping.