Abstract: The accurate remote sensing classification of surface coverage in reservoir areas provides important basic data for reservoir ecological environmental protection, reservoir optimization and dispatching, and soil and water conservation. In recent years, deep learning has been widely used in surface coverage classification. The current surface coverage classification method based on convolutional neural networks is usually a local spatial operation using regular convolutional kernels, which cannot easily capture long-distance dependencies between cells. Therefore, this paper proposes a dual Graph Convolutional Fusion Network（Dual-GCFN） for remote sensing classification of surface coverage. The network is divided into two parts. The upper branch uses the K-neighbor algorithm to construct a topological diagram between cells, and then uses the Graph Neural Network（GNN） to aggregate neighborhood features to update the target node information to achieve the aggregation of spectral features; the lower branch uses a dense Convolutional Network（DenseNet） to convolve local blocks of images to extract spatial information and high-order semantic information of remote sensing images. Finally, the spectral characteristics and spatial characteristics will be effectively fused through the dimensional channel Add operation, and after the fully connected layer, the Softmax classifier is used to realize the surface coverage classification. The Chushandian Reservoir area in Henan Province is the research object, and it uses Sentinel-2 remote sensing image data and DEM data in the research area to conduct experiments. The results show that the overall accuracy and Kappa coefficient of the Dual-GCFN classification proposed in this paper are 90.63% and 0.87, respectively. Compared with the classical classification method Random Forest classification（RF）, single GNN or DenseNet classification results, the overall accuracy increases by 15.97 percentage points, 10.86 percentage points, and 4.65 percentage points, respectively, the Kappa coefficient increases by 0.26, 0.24, and 0.08, respectively, and the misclassification of ground objects and the point noise in the classification map are greatly reduced in the classification map.