Multi-Branch Deep Learning Architecture for Satellite Image Classification using ViTNet-FCN with Transfer Learning

Satellite image classification plays a vital role in land cover mapping, resource management, and environmental monitoring. While recent advances in deep learning have improved performance, single model approaches such as CNNs and Vision Transformers often fall short in capturing the full complexity of satellite imagery. This work proposes a multi-branch fusion framework that integrates ResNet18, EfficientNetB0, a Fully Convolutional Network (FCN), and a Vision Transformer (ViT) within a unified architecture. We refer to the architecture as ViTNet-FCN. The proposed novel fusion strategy combines localized texture features from CNNs with the global spatial dependencies captured by Transformers, thereby overcoming the limitations of dual stream or ensemble methods. Extensive experiments on the EuroSAT dataset demonstrate that the framework outperforms strong baselines—including ResNet18 (93.4%), EfficientNetB0 (94.1%), FCN (91.7%), and ViT (95.0%)—achieving 98.76% accuracy, a macro F1 score of 0.9876, and consistently higher precision, recall, and Cohen’s kappa values. Beyond significant gains over individual models, the framework shows competitive performance against recent state-of-the-art CNN-Transformer hybrids, highlighting its robustness and generalization ability for satellite image classification. Although computationally more intensive than single models, the design establishes a scalable foundation for lightweight or pruned variants in future work.