Hybrid CNN–Transformer Architecture for Semantic Segmentation of High-Resolution Satellite Images

Abstract

Semantic segmentation of high-resolution satellite imagery is a crucial task for applications such as urban planning and environmental monitoring. However mod- ern satellite images are very complicated, containing objects of different sizes and irregular boundaries. Convolutional Neural Networks (CNNs) are powerful tools for extracting local details such as textures and edges but they struggle to capture long- range contextual information. On the other hand, Vision Transformers (ViTs) are good at context capturing on the global level but lack local precision and typically require massive amounts of training data. The project is intended to address this gap by developing a hybrid CNN-Transformer architecture. The network combines the strengths of CNNs for local feature ex- traction with the power of Transformers for global context understanding, leading to improved segmentation performance, particularly for large objects and complex landscape classes. The proposed hybrid model is designed, implemented, and tested on annotated satellite datasets, with performance measured using pixel accuracy and the Dice coefficient. The results show that the hybrid framework, especially with transfer learning, provides a significant improvement in edge detection and generalizes better in different environments. This work shows the importance of AI in remote sensing and offers a solid tool for automated decision-making systems.