As reported by the World Health Organization (WHO), depression will be the first mental disorder by 2030. Electroencephalogram (EEG) modality with deep learning has been adopted widely to mine the depressed patterns. However, the most of methods explore full-channel (the whole brain region) to extract the discriminative features for depression recognition, which resulting in some difficulties in the practical applications. To obstacle the challenges, we adopt several channels (represented as few-channel) to build a novel architecture, termed spatiotemporal characteristics with NeXtVLAD (STC-ND), for EEG-based depression recognition. Specifically, spatial features are learned by the Graph Convolutional Networks (GCN). Temporal features are learned by the multi-scale 1-D convolutional neural networks (1D-CNN), which can capture multi-scale dynamics from EEG. Then NeXtVLAD is adopted to aggregate the potential patterns from the spatial and temporal features. To validate the effectiveness of the proposed method, we carry out a large number of experiments on two public datasets, MODMA and EDRA, and the experimental results show that the accuracy of full-channel data reaches 99.48% and 99.13% respectively, while the accuracy of few-channel data reaches 98.96% and 96.26% respectively.