In the event of a single line-to-ground fault in a non-effectively grounded distribution network, reliable detection and isolation of faulty feeders aid in ensuring safe network operation. To improve the accuracy and reliability of faulty-feeder detection, this paper proposes a faulty-feeder-detection method based on the entire-frequency-domain fault characteristics. First, the zero-sequence current variation characteristics of feeders are analyzed using a wavelet packet algorithm; subsequently, the frequency-spectrum energy and zero-sequence current direction are employed to build a detection criterion. Second, the frequency-spectrum energy and direction criteria are combined to obtain a comprehensive detection criterion across the entire frequency spectrum. Third, the ensemble-learning algorithm is utilized to construct the proposed comprehensive criterion. An extreme-gradient-boosting algorithm is employed for efficient modeling along with large amounts of simulated data that are used as training dataset. To verify the reliability and generalization capability of the proposed method, PSCAD simulation, RTDS, and practical fault data are employed considering different topologies, parameters, and fault conditions of various distribution networks. The results obtained in this study reveal the proposed method to improve the accuracy of faulty-feeder detection significantly relative to conventional approaches, which demonstrates that the proposed method shows considerable reliability and generalization potential for practical utility.