Abstract: Quick and accurate landslide detection is vital for minimizing infrastructure damage and protecting the safety of human life. However, traditional landslide detection methods are time-consuming, labor-intensive, and inefficient; deep learning techniques based on satellite remote sensing imagery face challenges such as low training efficiency, demand for massive training samples, and inadequate identification of global features. Hence, this study proposed a novel landslide detection method based on the deep graph convolutional neural network (DGCNN). First, the graph-based image structure was utilized instead of image pixels to enhance model efficiency and the modeling capability for local features. Second, multi-dimensional image features, including the location, spectral indices, texture, and shape, were incorporated into the feature vector, reducing the model's reliance on massive training samples. Third, deep mutual-information evaluation functions were employed to enhance the global feature identification capability. Finally, the DGCNN-based landslide detection method was verified using two Sentinel-2 datasets, which were acquired from Brazil and China's Sichuan Province, respectively. The experimental results demonstrate that the DGCNN achieved an accuracy of above 80% under conditions of about 0.5% training samples and 200 training epochs, verifying its superiority and practicality.