Abstract: ‍ ‍In recent years， low-altitude UAV aerial images and high-altitude UAV remote sensing images have played an important role in land resource management， environmental monitoring， terrain mapping， urban planning， etc.How to accurately segment multiple categories in low-altitude UAV aerial images and high-altitude UAV remote sensing images has become one of the current research hotspots.For the low-altitude UAV aerial imagery and UAV remote sensing image segmentation tasks， it is difficult to capture long-distance spatial dependencies and difficult to extract dynamic connections between different batches of input categories，this paper proposes a semantic segmentation model GRNet based on ResNet and a modified graph convolution （GCN） module. Improved graph convolution takes the features extracted by ResNet as input， converts pixels into graph nodes， uses the affinity relationship between pixels as edges， and uses coordinate graph convolution and channel graph convolution to model global information for features， and at the same time， the category dynamic graph convolution is introduced， and the dynamic affinity of the input information is used as the edge to model the category dynamic correlation of the input features， and finally the extracted features are fused into the segmentation feature map to improve the segmentation accuracy of the model. This paper uses UDD5， UDD6 two low-altitude UAV aerial photography datasets and ISPRS Potsdam high-altitude UAV remote sensing dataset for training and verification， and in the experimental part respectively proves the role of different modules of this paper to improve graph convolution.The experimental results show that the proposed GRNet has good generalization performance， showing good segmentation performance in both the UAV aerial image data set and the UAV remote sensing image data set， which is better than multiple comparison networks.