Hybrid deep learning for satellite image classification: Integrating CNN and transformer architectures

Satellite image classification plays a critical role in modern remote sensing applications, including environmental monitoring, urban planning, agricultural assessment, disaster management, and intelligent Earth observation systems. However, accurate land-cover classification remains challenging due to inter-class similarity, intra-class variability, complex spatial patterns, and heterogeneous scene distributions. Conventional convolutional neural network (CNN)-based approaches demonstrate strong local feature extraction capability but often struggle to capture long-range contextual dependencies required for robust scene understanding. Conversely, Transformer-based architectures effectively model global contextual relationships but typically involve high computational complexity and large parameter counts. To address these limitations, this paper proposes a hybrid CNN–Transformer framework for satellite image classification that combines convolutional local spatial representation learning with Transformer-based global contextual modeling. The proposed architecture integrates a CNN backbone for extracting fine-grained spatial features, a Transformer encoder for capturing long-range semantic dependencies through self-attention mechanisms, and a feature fusion strategy for generating more discriminative and context-aware feature representations. In addition, an attention module is incorporated to emphasize informative spatial regions and improve robustness against intra-class variability. The proposed framework was evaluated on the EuroSAT dataset containing 27,144 satellite image samples across ten land-cover categories. Experimental results demonstrate that the proposed model achieved an overall classification accuracy of 98.36%, significantly outperforming the baseline CNN model and several state-of-the-art CNN-based, Transformer-based, and hybrid architectures. Comprehensive evaluation using Precision, Recall, F1-score, ablation analysis, class-wise performance assessment, and computational complexity analysis confirms the effectiveness and robustness of the proposed framework. Furthermore, the proposed architecture demonstrates an effective balance between classification performance and computational efficiency by maintaining moderate parameter count, inference time, and memory consumption compared with existing Transformer-based approaches. The obtained results confirm that integrating local feature extraction with global contextual modeling provides a highly effective solution for addressing the inherent challenges of remote sensing image classification.