ProX-Net: Orthogonal Projection Meets Proximal Topology for Efficient and Robust Remote Sensing Image Interpretation

Precise remote sensing image interpretation, particularly the extraction of key man-made structures like roads and buildings, is critical for intelligent transportation and urban planning. Existing methods struggle to balance global context with local details, often bottlenecked by the quadratic complexity and noise sensitivity of standard Transformers. To address these overlapping limitations, we propose ProX-Net, a novel architecture tailored for efficient semantic segmentation, driven by a hierarchical collaborative strategy of dual attention. Specifically, we introduce two core modules: the Orthogonal Projected Attention (OPA) features a linear reconstruction of cosine similarity to replace standard Softmax, effectively suppressing amplitude-sensitive background noise and ensuring global semantic consistency. Concurrently, the Proximal Topology Attention (PTA) employs a local topological sharpening mechanism to strictly preserve the connectivity of linear roads and the precise boundaries of block-like buildings. Extensive experiments on three diverse benchmarks (DeepGlobe, Massachusetts Roads, and WHU Buildings) demonstrate that ProX-Net consistently outperforms state-of-the-art methods, achieving an IoU of 79.59% and an F1-score of 86.61% on the DeepGlobe dataset. Furthermore, integrating these modules into the YOLO framework on the VisDrone dataset significantly boosts object detection, proving ProX-Net's versatile and practical application value across diverse remote sensing interpretation tasks.