Abstract: To address the challenge of off-road path planning within complex scenarios, this study proposed a path planning method for high-resolution remote sensing images integrating the HEDLink-RoadNet neural network and adaptive A* algorithm. By doing so, it aims to eliminate reliance on vector road network data inherent to conventional approaches and achieve end-to-end direct path planning. First, by synergizing the advantages of D-LinkNet and holistically-nested edge detection (HED) networks, this study designed the HEDLink-RoadNet model. This model then incorporated multi-scale side-output layers and a novel loss function with an inter-class balance coefficient, effectively mitigating the imbalance between road and non-road pixels and enhancing road extraction accuracy. Second, this study developed an adaptive A* algorithm. The algorithm can resolve dimensional mismatch in raster data-based path planning by dynamically adjusting traversal difficulty weights and dimensional distance in heuristic functions. Finally, this study further optimized search efficiency and minimized computational resource consumption using an upscaling strategy. Experimental results demonstrate that the proposed method achieved a path planning accuracy of 90.46% in complex scenarios, with significantly improved algorithmic efficiency. Compared to conventional approaches, the method streamlines the process while improving both efficiency and precision by eliminating road vectorization and network updating procedures. This method provides effective and robust technical support for disaster emergency response, geological survey route analysis, and intelligent transportation planning.