A Road Extraction Method Based on Dual-Domain Feature Fusion and Multi-Stage Fine-Tuned SAM

Road extraction from remote sensing imagery holds significant importance for smart city development, autonomous driving, and disaster emergency response. However, existing methods still exhibit limitations in modeling long-range dependencies and adapting to complex scenarios. To address these challenges, this paper proposes a Fourier-based Multi-stage Adaptive Fine-tuning framework (FM-SAM) that enhances road extraction performance through spatial-frequency dual-domain feature fusion. The proposed framework introduces three key innovations: (1) incorporation of Fast Fourier Transform (FFT) into the Segment Anything Model (SAM) to establish a frequency-domain attention mechanism for enhanced road edge feature representation; (2) development of an FFTLoRA fine-tuning adapter employing low-rank matrix decomposition for parameter-efficient optimization; and (3) implementation of a multi-stage adaptive fine-tuning strategy to balance model generalization and task-specific adaptation. Experimental results on both our self-constructed HF road dataset and the public Massachusetts dataset demonstrate that FM-SAM achieves IoU scores of 71.89% and 55.13% respectively, outperforming state-of-the-art models by 1.68% and 2.59%. Ablation studies further validate the effectiveness of both frequency-domain features and low-rank fine-tuning. This research presents a novel approach to road extraction that demonstrates superior performance potential in complex scenarios.