ResFormer: An Efficient Transformer Framework for Scalable Semantic Segmentation of Remote Sensing Images

Yanglin Ou^1,2Xiqi Wang^3*

• ¹Huizhou Technician Institute, Huizhou, China
• ²Huazhong University of Science and Technology, Wuhan, China
• ³Shandong Jianzhu University, Jinan, China

The translation of machine learning theory into operational remote sensing applications that deliver measurable societal value remains a paramount challenge. This endeavor requires models that are not only accurate but also scalable, reliable, and directly applicable to real-world problems such as climate resilience and sustainable urban development. While Convolutional Neural Networks (CNNs) have been foundational, their limited receptive fields often fail to capture the global context essential for interpreting complex scenes. Vision Transformers, with their global self-attention mechanism, offer a powerful alternative but typically incur prohibitive computational costs. To address these challenges, this paper introduces ResFormer, a novel architecture designed to bridge the gap between algorithmic innovation and demonstrable public good. Specifically, we propose a novel linear-complexity Transformer block integrated with residual connections, which drastically reduces the computational overhead from quadratic to linear complexity without sacrificing global context modeling. This efficiency enables the processing of high-resolution remote sensing imagery on commodity hardware. On the large-scale UAVid urban-scene dataset, ResFormer achieves a mean Intersection-over-Union (mIoU) of 68.7%, and on the ISPRS Potsdam dataset, it attains 85.9% mIoU. By holistically addressing scalability, reliability, and impact, ResFormer serves as a reproducible exemplar that moves the field toward machine learning systems that generate trustworthy and actionable knowledge for the public good. The implementation will be made publicly available to foster further research and application.