To address the issue of low reliability and high sample demand of data-driven high-impedance fault detection, a high-impedance fault identification method for distribution networks based on deep feature fusion in small sample conditions is proposed. First, based on the analysis of high-impedance fault characteristics in distribution networks, the waveform of high-impedance fault signals is deeply mined using discrete wavelet transform, and a multi-scale time-frequency feature graph and global statistical feature matrix are constructed to enhance the expression of high- impedance fault characteristics. Second, combining a lightweight residual network structure and self-attention mechanism, a deep feature extraction network is designed to realize the fusion extraction of local and global time-frequency features. Then, the metric meta-learning is introduced to calculate the feature class prototype and the distance between the class prototype and the sample in the metric space, so as to realize the construction of a high-impedance fault classifier. Finally, the effectiveness of the proposed method is verified by case studies with different complex operating conditions and field data sets.