<p>Next point-of-interest (POI) recommendation focuses on predicting the next location a user may visit by leveraging spatio-temporal patterns derived from their historical behaviors. However, accurately predicting the next POI is challenging due to data sparsity. While advances have been made in leveraging spatio-temporal information, existing methods often treat diverse features in isolation, fail to fully exploit multi-modal content, and lack the ability to dynamically adapt to user preferences. To address these challenges, this study proposes a novel framework integrating spatio-temporal correlation and multi-modal preferences (ISTMM) for next POI recommendation. First, a new multi-modal knowledge graph (MMKG) is constructed to integrate spatio-temporal information and multi-modal content, and pre-trained models are utilized to overcome the semantic discrepancies between different modalities. Second, a similarity function is designed to construct similarity matrices based on the learned multi-modal features and spatio-temporal features, and enrich the representation of POIs through graph convolutional network (GCN). Third, a new preference-weighting strategy is proposed to better aggregate the user's historical behaviors with the contextual information by modeling spatio-temporal correlation and multi-modal preferences. The performance and stability of ISTMM in achieving higher accuracy in next POI recommendation are validated through comprehensive evaluations across four real-world datasets.</p>