To solve the problem of performance degradation caused by insufficient array signal feature representation under limited snapshot conditions, this research proposes a novel deep transfer learning approach specifically tailored for direction-of-arrival (DOA) estimation. Our approach synergistically combines the residual attention network (RAN) with the bidirectional long short-term memory network (BiLSTM) to construct a robust pretrained model that efficiently captures and extracts abundant features from complex signal data. The model is initially pretrained on a large dataset, allowing it to learn rich feature representations that are later fine-tuned for deployment in scenarios with restricted snapshots. Through this careful fine-tuning process, the pretrained knowledge is effectively transferred, resulting in significantly optimized model parameters and improved estimation performance. Experimental findings clearly show that our proposed method not only considerably improves the DOA estimation accuracy but also maintains excellent robustness under challenging snapshot-restricted conditions, providing a promising new solution for real-time array prediction in environments with limited data availability.