耕地作为国家粮食生产的重要保障，其空间分布是粮食安全评估、土地资源管理等领域的主要依据，耕地信息提取在农田保护决策和作物生产规划中被广泛需求。现有的耕地信息提取方法忽视了地块的差异化特征和边缘细节蕴含的丰富信息，且提取结果碎片化，边界模糊。基于以上问题，论文提出了一种结合EfficientNet骨干网络和U型框架构建的改进型耕地信息提取模型BECU-Net（Boundary Enhancement Classification U-Net），该方法针对边缘欠拟合问题，为实现边缘特征和深度特征的信息互补，设计由CoT模块（Contextual Transformer Module）、门控卷积、scSE（Spatial-Channel Sequeeze and Excitation）注意力机制形成的边缘分支子网络，并以此提高模型处理边界信息的专注度。同时，构建含约束项的联合型边缘增强损失函数BE-LOSS（Boundary Enhancement Loss）进一步完善模型运算性能。研究使用GID高分2号RGB-NIR四波段数据，与梯度、指数、纹理特征图共同构建耕地特征机制。实验分别与不同网络结构、不同损失函数的模型进行对比。结果表明：改进算法的总体精度和F1分数均有改善，相比于DeeplabV3+网络，提取精度提升2.24%，F1分数提升1.77%。综上，论文算法为进一步解决耕地信息提取时边界模糊问题提供技术参考，为复杂交界的精准划分提供理论支撑。

As an important guarantee of national food production, the spatial distribution of cultivated land is an important basis for food security assessment, land resource management and other fields. The extraction of cultivated land information is widely needed in farmland protection decision-making and crop production planning. The existing cultivated land information extraction methods ignore the land parcel differentiation features and the rich information contained in the edge details, and the extraction results are fragmented and the boundary is fuzzy. Based on the above problems, this paper proposes an improved cultivated land in-formation extraction model BECU-Net (Boundary Enhancement Classification U-Net), which is constructed by combining efficientnet backbone network and U-shaped framework. Aiming at the problem of edge un-der fitting, in order to realize the information complementarity of edge features and depth features, this method designs an edge branch sub network formed by CoT (Contextual Transformer Module), gated con-volution and SCSE (Spatial Channel Sequence and Exception) attention mechanism, so as to improve the attention of the model in processing boundary information. At the same time, the joint edge enhancement loss function BE-Loss (Boundary Enhancement Loss) with constraints is constructed to further improve the operation performance of the model. The study uses the RGB-NIR four band data of GID Gaofen No. 2, to-gether with the index and texture feature map, to construct the cultivated land feature mechanism. The ex-perimental design is to compare Our model with different network structures and different loss functions. The experimental results show that both the overall accuracy and F1 score of the improved algorithm are im-proved. Compared with DeeplabV3+ network, the extraction accuracy is improved by 2.24% and F1 score is improved by 1.77%. In conclusion, the algorithm in this paper provides a technical reference for further solv-ing the boundary fuzzy problem in farmland information extraction, and provides a theoretical support for the accurate division of complex boundaries.