Abstract: Due to the combined effects of complex spectral mixtures, blurred boundaries of ground objects, and environmental noise, it is extremely challenging to accurately identify water boundaries from high-resolution remote sensing images. To address this problem, proposes a water body extraction method based on ConvNeXt and dual feature extraction branch (CoNFM-Net) on the basis of PSPNet. In the encoder stage, ConvNeXt is used instead of Resnet50 as the backbone network, which uses inverted bottleneck layer, large kernel and other designs to enhance the feature extraction ability of the network. In the decoder stage, a dual feature extraction branch structure with multi-scale feature fusion and context information enhancement is designed. In order to effectively utilize the multi-level feature map generated by the backbone network, a bidirectional feature fusion module (BiFMM) is designed to solve the problem of scale inconsistency in boundary recognition. Aim to improve the utilization rate of global information, the deep feature map output by the backbone network is passed through the global context information module (GCIM). At the same time, the deepest feature map of the multi-scale feature fusion branch is spliced with it to enhance the model 's ability to capture the details of the water boundary. The experimental results show that the mean intersection over union and F1-Score of this method on LoveDA dataset, GF-2 dataset and Sential-2 dataset are 89.64 %, 94.32 %, 92.60 %, 96.16 % and 93.72 %, 96.73 %, respectively. In the same environment, compared with UNet, DANet, CMTFNet and other semantic segmentation algorithms, the proposed algorithm CoNFM-Net has certain advantages.

Key words: water body extraction, ConvNeXt, high-resolution remote sensing images, feature fusion, dual feature extraction branch

摘要： 由于复杂的光谱混合物、地物边界模糊、环境噪声等因素的共同作用，从高分辨率遥感图像中准确识别水体边界极具有挑战性。针对此问题，在PSPNet的基础上提出基于ConvNeXt与双特征提取分支的水体提取方法（CoNFM-Net）。在编码器阶段，以ConvNeXt代替Resnet50作为主干网络，利用逆瓶颈层、大卷积核等设计来增强网络的特征提取能力。在解码器阶段，设计了多尺度特征融合和上下文信息增强的双特征提取分支结构，多尺度特征融合分支为有效利用主干网络产生的多层次特征图，设计了一种双向特征融合模块（BiFMM），以解决边界识别中尺度不一致的问题；上下文信息增强分支为提高全局信息的利用率，将主干网络输出的深层特征图通过全局上下文信息获取模块（GCIM）。同时，将经过多尺度特征融合分支的最深层特征图与其进行拼接，增强模型对水体边界细节的捕捉能力。实验结果表明，该方法在LoveDA数据集、高分二号(GF-2)数据集及Sential-2数据集上的平均交并比和F1分数分别为89.64%、94.32%，92.60%、96.16%及93.72%、96.73%，且在同样环境下，与UNet、DANet、CMTFNet等语义分割算法相比，该算法CoNFM-Net具有一定优势。

关键词: 水体提取；ConvNeXt；高分辨率遥感影像；特征融合, 双特征提取分支结构