针对滨海湿地植被光谱特征相似而易被混淆分类的问题,本文提出了结合深度学习和植被指数的滨海湿地信息提取网络MFVNet。该网络以高分辨率遥感影像特征和典型植被指数为输入,将UNet中的双卷积操作替换为本文提出的增强多尺度特征提取模块,用于捕获不同尺度的上下文特征,并在解码器中将不同分辨率的语义特征图和细节特征图进行融合,增强了滨海湿地地物的特征表示。在黄河口滨海湿地高分二号遥感影像上进行了实验,结果表明：(1)深度学习方法的信息提取精度普遍优于传统的机器学习分类方法SVM；相比HRNet等深度语义分割网络,MFVNet在滨海湿地所有植被类型上都能取得更好的信息提取结果；(2)将修正土壤调节植被指数MSAVI、差值植被指数DVI和比值植被指数RVI与高分二号影像拼接对滨海湿地信息提取贡献较大。

Objective: The biomass and growth of coastal wetland vegetation vary greatly due to the different water and salt conditions in the growing area, and the spectral features of certain vegetation at the peak biomass are very similar, which makes coastal wetland vegetation easy to be misclassified. In response to this problem, this paper proposes a new semantic segmentation network MFVNet to be combined with vegetation index for fine mapping of coastal wetlands. Method: In the proposed MFVNet, an enhanced multi-scale feature extraction (E-MFE) module is first constructed based on atrous convolution and attention mechanism to adaptively capture features of different scales. Then, the E-MFE module was used to replace the double convolution operations in traditional encoder-decoder, such as UNet. It was also used to merge the semantic features and detailed features of different resolutions to enhance feature representation. Finally, some typical vegetation indices were selected and input into the proposed MFVNet to improve the ability of coastal wetlands fine mapping. Result: The experiments of this paper were conducted using GF-2 remote sensing images to study the coastal wetlands of the Yellow River Estuary. Experimental results indicated that the proposed MFVNet achieved good performance with an overall accuracy of 93.89% and a Kappa of 0.9072. On typical vegetation, such as reeds, spartina alterniflora, tamarix mixed area, and seagrass beds in the Yellow River Estuary, the F1 scores of MFVNet were 0.91, 0.87, 0.82, and 0.76, respectively, which were better than other methods. At the same time, ablation experiments showed that the combination of the E-MFE module and the vegetation index can increase the overall accuracy from 91.46% to 93.89%. Conclusion: (1) Compared with deep semantic segmentation networks such as HRNet, MFVNet can more effectively extract vegetation information of coastal wetlands; (2) The proposed EMFE module can adaptively capture features of different scales and improve the overall accuracy, which justified its effectiveness in coastal wetland mapping; (3) The inclusion of vegetation index can enhance the spectral features of coastal wetland vegetation and improve the accuracy of vegetation information extraction, which indicated the importance of vegetation index in coastal wetland mapping; (4) Simultaneously splicing modified soil adjusted vegetation index (MSAVI), difference vegetation index (DVI), and ratio vegetation index (RVI) in remote sensing images contributed the most to the extraction of coastal wetland information.