Automated field delineation is a challenging task due to variations within pixel values and limited differences between classes. This study proposes an effective strategy to tackle this issue. Firstly, a framework combining the Canny operator and Watershed segmentation algorithm (CW) quickly labels the training dataset, reducing manual vectorization workload. Secondly, a CW-trained recurrent residual U-Net deep semantic segmentation network mines both low-level and deep semantic features. Finally, a boundary connecting method integrates fragmented boundaries to generate the agricultural field boundary.