With the rapid and accurate acquisition of rural road information, essential data are provided for agricultural machinery operation navigation and high-standard farmland construction evaluation. To address challenges such as occlusion, small spectral differences, and diverse geometric shapes in complex rural environments, an improved semantic segmentation model, SMC_ResUnet, was proposed based on the Res-Unet architecture with controllable encoding depth. Using ResUnet50 as the backbone, the strip pooling module was introduced in the encoder to enhance the extraction of long-range spatial features of rural roads. Additionally, the CA attention module was incorporated into the residual blocks to improve the perception of subtle road features through positional information, thereby reducing omission errors. A hybrid pooling module was integrated into the encoder-decoder pathway, combining strip pooling and standard pyramid pooling to balance the recognition of rural roads with diverse shapes while minimizing false positives. The proposed model was validated on a high-resolution rural road dataset from Nenjiang City, Heilongjiang Province. Experimental results demonstrated that SMC_ResUnet outperformed comparison models, achieving an average accuracy of 98.58%, recall of 83.40%, MIoU of 78.06%, and F1-score of 85.89%, with an overall accuracy of 97.41% in large-scale rural road extraction. Ablation experiments confirmed the effectiveness of each module in addressing specific challenges of rural road identification. The model's generalization capability was further verified by using the Deep Globe Road Extraction Dataset. The research result can provide a valuable reference for acquiring rural road information and guiding agricultural machinery navigation.