【摘要】 为了准确提取水稻的种植面积，提出一种基于自适应特征提取网络模型RC＿UNet和特征组合的水稻面积提取方法：(1)设计一种自适应通道卷积注意力机制(ACCA),通过动态调整水稻特征通道权重，实现RC＿UNet对于水稻特征的自适应提取；(2)构建一种特征对齐模块(FAM),通过特征对齐，来减小特征之间的差异性，促进特征融合，增强网络的特征表达能力；(3)构建了4种用于水稻面积提取的特征组合方案。以2021年湖北省石首市Sentinel-2影像为数据源，进行实验。实验结果表明，与SE、CBAM相比，ACCA能够显著提升网络收敛效果，提升网络收敛效果。同时验证了，近红外(NIR)、主成分第一分量(PCA1)、增强型植被指数(EVI)特征组合更加适用于提取水稻区域。在此基础上，RC＿UNet网络模型的F1分数、交并比、相对面积误差分别为89.21%、81.52%、4.96%,优于SegNet、DeepLabV3+和U-Net网络模型。更多还原

【Abstract】 In order to accurately extract rice planting area, a method of rice area extraction based on adaptive feature extraction network model RC＿UNet and feature combination was proposed. Firstly, an adaptive channel convolutional attention mechanism(ACCA) was designed to dynamically adjust the weight of rice feature channels to realize the adaptive extraction of rice features by RC＿UNet. Secondly, a feature alignment module(FAM) was constructed to reduce the differences among features, promote feature fusion, and enhance the feature expression ability of the network. Finally, four feature combinations for rice area extraction were constructed. The Sentinel-2 images of Shishou City, Hubei Province in 2021 were used as data sources to conduct experiments. The experimental results show that compared with SE and CBAM,ACCA can significantly improve the effect of network convergence. At the same time, it was verified that the feature combination of Near Infrared(NIR),Principal Component Ⅰ(PCA1) and Enhanced Vegetation Index(EVI) was more suitable for extracting rice regions. On this basis, the F1 score, intersection over union and relative area errors of RC＿UNet network model are 89.21%,81.52% and 4.96%,respectively, which are better than SegNet, DeepLabV3+ and U-Net network models.更多还原