High-resolution remote sensing image road extraction system: data-model collaborative optimization

Background: Accelerating global urbanization and increasing transportation demands highlight the critical need for precise and efficient road system planning in geospatial applications. Existing methods for extracting roads from high-resolution remote sensing images face challenges in balancing accuracy and computational efficiency, especially in complex urban environments. To address this, we introduce a data-model co-optimization framework that enhances both accuracy and efficiency. Methods: Our research focuses on three key aspects: (1) Multi-dimensional data quality optimization through a fourth-order image processing chain ("Gaussian noise reduction - non-local mean denoising - CLAHE contrast enhancement - edge enhancement") combined with mask double median filtering and morphological closure operations to improve input data quality. (2) Development of a Lightweight Dynamic Attention Network Architecture (MACL-U-ResNet) based on ResNet50, integrating four attention modules (SE/CBAM/ECA/CA) to dynamically select channel-space-position perceptions via hyperparameter φ . This architecture includes a depth-feature retention structure and cross-layer fusion decoder. (3) Proposal of an efficiency-accuracy balance coefficient η (η = IoU/parameter count) to evaluate resource-constrained scenarios, revealing the gains of data preprocessing on system efficiency. Results: Experiments on the DeepGlobe dataset show that our system achieves 66.82% IoU and 79.32% F1-Score, surpassing the state-of-the-art LCMorph model by 0.52% in IoU while reducing the parameter count by 39.9%. The η coefficient value of 2.33 validates the system's balanced accuracy-efficiency, offering a significant advantage for edge computing applications.This study demonstrates a robust solution for road extraction, advancing the field towards more efficient and accurate geospatial applications.