Abstract: Extracting water bodies from high-resolution remote sensing images plays a crucial role in flood monitoring. Current techniques mainly utilize deep convolutional neural networks (DCNNs) to extract water regions from remote sensing images. However, remote sensing images contain rich information and exhibit characteristics of “high intra-class variation and low inter-class variation”, leading to phenomena such as missed detections, false detections and blurry identification in water body recognition using DCNNs. This paper proposes a novel dual path learning network (DPL-Net) for accurately and efficiently extracting water body information. The shallow water features learned by the feature extraction module are input separately into the semantic path and spatial path. The high-level semantic information learned from the semantic path is fused with the spatial characteristics learned from the spatial path to achieve complete water body extraction. This network introduces a residual semantic learning module (RSL) in the semantic path to learn more discriminative semantic information, and a multiscale residual expansion convolution space path (MRE) in the spatial path to encode rich multi-scale contextual information. To make the network training more focused on water regions, this paper incorporates intersection over union (IOU) as a component of the loss function. Experimental results show that compared to existing advanced methods, the proposed algorithm achieves an IOU of 89.56% on the public GID dataset, outperforming other networks. To validate the generalization and robustness of the model, experiments are conducted on the LoveDA dataset, showing an IOU of 73.77%, which is higher than other models.

Key words: flood monitoring, remote sensing image, convolutional neural network, semantic segmentation, water body extraction

摘要： 提取高分辨率遥感图像中的水体，对洪涝监测具有至关重要的作用。现有技术主要利用深度卷积神经网络（deep convolutional neural network，DCNN）提取遥感图像中水体区域。然而，遥感图像蕴含丰富信息，表现出“类内差异大，类外差异小”这一特征，导致DCNN对水体识别提取出现“漏判”“误判”及“识别模糊”等现象。提出一种新颖的双路径学习网络（dual path learning network，DPL-Net）用于准确高效地提取水体信息。将特征提取模块学习的浅层水体特征分别输入语义路径与空间路径，将语义路径学习的高层语义信息与空间路径学习的空间特性进行特征融合，起到提取完整水体的作用。该网络在语义路径中提出残差语义学习模块（residual semantic learning module，RSL）学习更具判别特性的语义信息，在空间路径中提出多尺度残差扩展卷积空间学习模块（multiscale residual expansion convolution space path，MRE）编码丰富的多尺度上下文信息。为了使网络训练更加注重水体区域，将交并比（intersection over union，IOU）作为损失函数的组成部分。实验结果表明，与现有先进方法相比，所提算法在公开数据集GID上IOU指标为89.56%，优于其他网络。为了验证模型的泛化性和鲁棒性，在LoveDA数据集上进行实验分析，结果表明IOU指标为73.77%，高于其他模型。

关键词: 洪涝监测, 遥感图像, 卷积神经网络, 语义分割, 水体提取