基于双路图卷积的黄河三角洲湿地地物分类研究

黄河三角洲湿地地物精确分类对湿地资源的保护、开发和利用具有重要意义。目前的湿地分类算法大多存在着全局信息利用不足,地物类型边界不易区分等问题,导致分类精度不高。针对此问题,本文提出了基于双路图卷积的黄河三角洲湿地地物分类算法,包括图结构数据构建模块、特征提取与融合模块两部分。图结构数据构建模块,设计欧式图表示光谱值之间的绝对差异,衡量不同地物类型,设计余弦图表达不同像素光谱波形之间的差异,用以区分不同的地物边界;特征提取与融合模块,利用图卷积聚合全局信息,对欧式图利用双层图卷积进行特征提取,对余弦图使用图U-Net网络进行特征提取,之后将两个特征融合,得到同时具有光谱值绝对差异和光谱波形差异的融合特征,最后进行分类。在CHRIS和GF5两个数据集的实验结果表明,本文所提算法在黄河三角洲湿地地物分类中取得了具有竞争力的分类结果。

Yellow River Delta wetland classification based on a dual graph convolution network

Accurate wetland classification of the Yellow River Delta is crucial for the protection, development, and utilization of wetland resources. Most of the current wetland classification algorithms have limitations, such as insufficient use of global information and difficulty in distinguishing the boundary of ground object types. These drawbacks lead to low classification accuracy. To solve this problem, a dual graph convolution network is proposed for the Yellow River Delta wetland classification. This network includes modules for graph structure data construction and feature extraction and fusion. A Euclidean graph is designed to represent the absolute difference between spectral values and to measure different object types. A cosine graph is designed to express the difference between the spectral waveforms of different pixels to distinguish different object boundaries. In the feature extraction and fusion module, graph convolution is applied to aggregate global information. Two graph convolution layers are used to extract features from the Euclidean graph, and the graph U-Net is utilized to extract features from the cosine graph. Finally, the features extracted from the two graphs are fused to obtain features with absolute differences in spectral values and differences in spectral waveforms for classification. Experiments conducted on the CHRIS and GF5 datasets demonstrated the effectiveness of the proposed method for Yellow River Delta wetland classification.