Accurate and efficient extraction of farmland parcels is essential for high-resolution remote sensing applications in precision agricultural management. However, current segmentation methods frequently encounter difficulties in accurately delineating blurred boundaries and preserving fine-grained details, particularly within complex and fragmented farmland landscapes. To address these challenges, we propose EdgeGuided-Net, a dual-branch segmentation framework based on the Pyramid Vision Transformer (PVTv2), which integrates boundary guidance with multi-scale contextual learning. The Main Branch leverages a Boundary Guided Module (BGM) to adaptively fuse semantic features and boundary information, improving the network’s ability to perceive weak boundaries and complex parcel structures. Meanwhile, the Edge Branch introduces a novel Learnable Laplace and Depthwise Residual (LLDR) module, which improves the representation of high-frequency edges by integrating parameterized Laplacian kernels with depthwise separable residuals. A cross-branch fusion mechanism integrates precise boundary cues with rich semantic context, achieving significant improvements in both boundary sharpness and region consistency. Extensive experiments conducted on a variety of farmland scenes demonstrate that EdgeGuided-Net achieves superior performance over state-of-the-art methods in terms of segmentation accuracy and boundary delineation. This approach offers a robust and generalizable solution for farmland parcel extraction, providing technical support for critical agricultural remote sensing tasks. These include monitoring farmland fragmentation, detecting non-agricultural land conversion, and assessing grain production.