To address the strong stealthiness and low detectability of false data injection attacks (FDIAs) in DC microgrids, this paper proposes a FDIA detection method based on deep spatiotemporal feature learning. First, a parallel dual-branch detection model is constructed. One branch incorporates a Transformer module to extract global information and cross-node features through a self-attention mechanism, while the other branch adopts a gated recurrent unit (GRU) to capture temporal dependencies and dynamic evolution patterns in measurement data. Second, feature-scale alignment and adaptive weighting are employed to achieve feature-level fusion of spatial and temporal representations, supplemented by normalization and residual mechanisms to suppress redundancy and noise. Then, the fused features are fed into a neural network classifier to enable unified detection of multiple types of FDIA. Finally, a multi-type attack dataset is constructed under typical DC microgrid scenarios, and comparative experiments are conducted. The results demonstrate that the proposed method outperforms baseline models across overall evaluation metrics and exhibits strong robustness and generalization capability.