URSE-Net: a method of extraction winter wheat planting information based on multispectral remote sensing images

文献类型： 外文期刊

第一作者： Zhou, Yuguang

作者： Zhou, Yuguang;Ma, Mingxu;Yang, Guang;Wang, Zeyu;Lai, Youchun;Liu, Haifan;Chen, Li;Zhou, Yuguang;Ma, Mingxu;Yang, Guang;Wang, Zeyu;Lai, Youchun;Liu, Haifan;Chen, Li;Zhou, Yuguang;Ma, Mingxu;Yang, Guang;Wang, Zeyu;Lai, Youchun;Liu, Haifan;Chen, Li;Wang, Mo;Wang, Mo

作者机构：

关键词： URSE-Net; CNN; semantic segmentation; Image Recognition; information extraction

期刊名称：INTERNATIONAL JOURNAL OF REMOTE SENSING （ 影响因子：2.6； 五年影响因子：2.9 ）

ISSN： 0143-1161

年卷期： 2025 年 46 卷 16 期

页码：

收录情况： SCI

摘要： Deep learning techniques are increasingly applied to research on crop identification, area detection, yield prediction, etc. However, some current research results show that there are problems such as low accuracy and obvious overfitting when using remote sensing images for the above studies. Based on this condition, a model URSE-Net was proposed, which is based on the U-Net structure and incorporates Feature Pyramid Attention (FPA) with attention mechanisms to integrate multi-scale information, reduce background interference, and enhance recognition accuracy. Dropout layers are added after max-pooling and convolution operations to suppress overfitting, and the enhanced feature extraction network is structurally simplified to reduce computational load and model complexity. A 4-channel remote sensing image of Gaofen-2 was chosen as the data source for conducting a winter wheat planting information extraction experiment with the planting area in Xian County, Hebei Province, China. The results show that (i) the URSE-Net model performs best when the batch size is set to 8, with a pixel accuracy (PA) of 97.11% and mean Intersection over Union (mIoU) of 93.78%; (ii) the multispectral remote sensing images can improve the performance of the model compared with the RGB images; (iii) the method proposed in this study has greater PA and mIoU than the U-Net, Segnet, PSPnet, and Deeplabv3+ models; and the URSE-Net method has the best performance compared to other methods in extracting information for each category in the 3-category dataset. Experimental results demonstrate that the optimization methods proposed in this study effectively address issues such as low accuracy and overfitting present in existing models, providing valuable insights for future model improvements.

分类号：

  • 相关文献
作者其他论文 更多>>

微信小程序