无线电工程

利用深度学习实现遥感影像耕地区域自动化检测，取代人工解译，能有效提升耕地面积统计效率。针对目前存在分割目标尺度大且连续导致分割区域存在欠分割现象，边界区域情况复杂导致边缘分割困难等问题，提出了语义分割算法——Swin Transformer, TransFuse and U-Net (SF-Unet)。为强化网络不同层次特征提取和信息融合能力，提升边缘分割性能，使用U-Net网络替代TransFuse网络中的ResNet50模块；将Vision Transformer (ViT)替换为改进后的Swin Transformer网络，解决大区域的欠分割问题；通过注意力机制构建的Fusion融合模块将2个网络输出特征进行融合，增强模型对目标的语义表示，提高分割的精度。实验表明，SF-Unet语义分割网络在Gaofen Image Dataset (GID)数据集上的交并比(Intersection over Union, IoU)达到了90.57%,分别比U-Net和TransFuse网络提升了6.48%和6.09%,明显提升了耕地遥感影像分割的准确性。

Using deep learning to realize automatic detection of cultivated land areas in remote sensing images, replacing manual interpretation, can effectively improve the statistical efficiency of cultivated land areas. Aiming at the current problems that the segmentation target scale is large and continuous, which leads to under-segmentation in the segmented area, and the complexity of the boundary area makes edge segmentation difficult, the Swin Transformer, TransFuse and U-Net(SF-Unet) semantic segmentation algorithm is proposed. In order to strengthen the feature extraction and information fusion capabilities of different levels of the network and improve the edge segmentation performance, the U-Net network is used to replace the ResNet50 module in the TransFuse network; the Vision Transformer(ViT) is replaced by improved Swin Transformernetwork to solve the problem of under-segmentation of large areas; the Fusion module constructed by the attention mechanism fuses the output features of the two networks to enhance the semantic representation of the model to the target and improve the accuracy of segmentation. Experiments show that the Intersection over Union(IoU) ratio of the SF-Unet semantic segmentation network on the Gaofen Image Dataset(GID) has reached 90.57%, which is 6.48% and 6.09% higher than U-Net and TransFuse networks, respectively. The accuracy of cultivated land remote sensing image segmentation is significantly improved.