Abstract: Remote sensing images exhibit diverse scene categories with high inter-class similarity. Existing Convolutional Neural Network (CNN) classification methods often entail large parameter counts and encounter challenges in distinguishing similar features and focusing on key features. Therefore, a hybrid model combining a lightweight CNN and Transformer architecture is proposed. The CNN module employs partial convolutions to efficiently extract spatial features and integrates a coordinate attention mechanism to reduce computational overhead while capturing broader information., and utilizes random channel shuffling to enhance inter-channel information exchange and feature diversity. Within the Transformer module, stability during training is ensured by introducing the MobileVIT block, guarding against overfitting on low-level features. CNN handles low-level features, while the Transformer manages high-level features, creating a complementary synergy that enhances overall classification performance. Comparative experimental results on the publicly available datasets NWPU-RESISC45 and VGoogle demonstrate that the model achieved classification accuracies of 95.69% and 97.56%, respectively, with parameter and computational complexities of only 1.29 MB and 269.81 MB. This further validates its effectiveness.