摘要：

在光学卫星影像中, 表碛覆盖型冰川的光谱和山地、岩石极为相近, 导致冰川与周围地形难以有效区分, 使得冰川的自动化分割变得困难。针对这一问题, 提出一种基于光学卫星图像和数字高程模型(DEM)的双输入图像语义分割网络(DENet)。该网络采用双编码框架, 结合多尺度特征提取和注意力机制, 通过整合来自不同数据的特征信息, 获取DEM地貌参数, 以解决表碛覆盖型冰川中同谱异物导致的源头区域误分割问题。首先通过多尺度可分离卷积注意力模块和多核注意力池化模块对卫星图像和DEM分别进行特征提取, 然后将获取到的2个特征图进行融合。多尺度特征提取模块可用于捕捉和融合冰川图像的多个尺度信息, 以产生更丰富和全面的特征表示。同时, 引入注意力机制可以对每个通道和空间位置分配不同的权重, 关注不同尺度上的特定区域, 使模型能够聚焦于更重要的信息, 减少多余特征的影响。实验结果表明, 该网络的平均交并比(IoU)达到94.6%, 比U-Net、DeepLabv3+网络分别提高4.53和3.38百分点, 其能提升山地冰川区域的分割准确率。

关键词: 光学卫星影像, 双编码, 表碛, 语义分割, 数字高程模型

Abstract:

In optical satellite imagery, the spectra of debris-covered glaciers are remarkably similar to those of mountainous terrain and rocks. This poses challenges in distinguishing glaciers from the surrounding topography and makes automated segmentation difficult. To address this problem, a Dual-Encoding Network (DENet) based on optical satellite images and Digital Elevation Model (DEM) is proposed. The network employs a dual-encoding framework that integrates multi-scale feature extraction and attention mechanisms. By incorporating features from different data sources and extracting DEM topographic parameters, it addresses mis-segmentation issues in source areas caused by spectrally similar objects in debris-covered glaciers. First, the satellite image and DEM extract features using multi-scale separable convolution attention and multi-kernel attention pooling modules. Subsequently, the obtained two feature maps are fused. The multi-scale feature extraction module captures and integrates information from glacial images of various scales to generate more comprehensive and enriched feature representations. The attention mechanisms simultaneously assign different weights to each channel and spatial position, focusing on specific regions at different scales. This enables the model to concentrate on critical information and reduce the impact of redundant features. Experimental results demonstrate that the model achieves an average Intersection over Union (IoU) of 94.6%, surpassing those of the U-Net and DeepLabv3+ networks by 4.53 and 3.38 percentage points, respectively. This improvement enhances the accuracy of mountain glacier region segmentation and validates the competitiveness of the proposed network compared with other existing models.

Key words: optical satellite image, dual encoding, debris-cover, semantic segmentation, Digital Elevation Model (DEM)