Unintentional islanding events pose a significant threat to the stable operation of hybrid single- and three-phase hybrid microgrid clusters. Timely and accurate detection of the actual islanding status is essential to ensure their secure and reliable operation. However, traditional islanding detection methods suffer from detection blind spots, and many weakly correlated or irrelevant electrical characteristics can adversely affect the accuracy of islanding detection. To address these issues, an islanding detection method for hybrid single- and three-phase microgrids based on the improved adaptive boosting (Adaboost) method is proposed. First, the slime mould algorithm (SMA) is introduced into the Adaboost learner to improve classification ability and reduce disturbance effects. Additionally, the weight learning method using group switchable normalization (GSN) is adopted to shorten the detection time. Then, an islanding detection model is established using the proposed SMA-Adaboost learner. To further enhance the efficiency and accuracy of the islanding detection model, electrical features strongly correlated with islanding status are extracted based on the partial least squares (PLS) method. Finally, the performance of the proposed method is validated through simulations on a hybrid single- and three-phase microgrid cluster based on the improved IEEE37-bus system. The results demonstrate that the proposed islanding detection method can accurately detect islanding without being affected by disturbance signals or three-phase system unbalance, and exhibits superior accuracy and generalization capability compared to existing detection methods.