Optimization of region segmentation techniques in remote sensing image road extraction based on dual stage feature fusion

The continuous advancement of remote sensing technology has made the importance of Remote Sensing Image (RSI) processing increasingly prominent in fields such as urban planning and traffic management. To address low accuracy in road extraction from RSIs in complex environments, this study proposes a RSI region segmentation model based on dual-stage feature fusion. Compared to existing methods, this model introduces a small target activation module specifically designed for small targets. It dynamically focuses on and enhances weak target features through dual spatial and channel attention. A fuzzy refinement fusion module is designed to implement an adaptive two-stage feature fusion mechanism that can dynamically adjust the fusion weights of high-level semantics and low-level details according to the context. Comparative experiments show that on the aerial image object segmentation dataset, the overall accuracy reaches 98.4±0.5%, the average intersection to union ratio is 82.6±0.7%, the running time is 1.8±0.08 seconds, and the memory occupation is 2.5±0.12 GB, all of which are superior to other algorithms. The segmentation display for specific roads has a small target recognition accuracy of 95.7%, with tree and vehicle recognition rates of 85.2% and 85.4%. The accuracy of road edge segmentation is 84.3%, while the misidentification rates of buildings and shadows are only 3.1% and 2.8%, and it performs stably under different lighting and climate conditions. The above results indicate that the proposed segmentation model effectively improves the stability of road extraction, providing reliable technical support for urban planning and traffic management. • To address low accuracy in road extraction from RSIs in complex environments, this study proposes a RSI region segmentation model built on dual stage feature fusion. • This model deeply extracts image features through a dual attention network and performs segmentation processing on the image through dual stage feature fusion.