Abstract: In the context of precision agriculture, recent years have witnessed a remarkable increase in the abundance of remote sensing image resources, accompanied by a significant boost in computing power. This progress has led to more refined land cover classification maps through semantic segmentation of remote sensing images, which is of great significance for farmland ecosystem analysis. In the agricultural domain, analyzing the distribution characteristics of farmland and roads,key components of the rural infrastructure network, enables accurate determination of straw yield and the distribution data of farmland roads. This information is crucial for formulating scientific and reasonable collection, storage, and transportation route plans, which is essential for efficient agricultural resource management.In existing research, within the scope of agricultural informatics, the application of deep-learning models to extract farmland road information encounters challenges such as high complexity. In this study, leveraging semantic segmentation technology, a core technique in remote sensing data processing for agriculture, research on the extraction of farmland road information from remote sensing images is carried out. Based on the land cover classification extracted from the images, research on optimizing the location selection of straw collection and storage stations is conducted, which is an important part of agricultural waste management.Through ablation experiments, it is demonstrated that both the asymmetric fusion non - local block (AFNB) and the Structure of the Dual Attention Module contribute positively to the segmentation effect. Compared with the original network model, the integrated results after combining these two components lead to increases in IoU, Acc-road, and mIoU by 5.2, 7.78, and 2.73 percentage points, respectively. By utilizing class activation graph analysis, this model significantly enhances the accuracy and efficiency of farmland road information extraction, a fundamental task in agricultural remote sensing. Its superiority is validated across multiple datasets, with achieving a minimum mIoU of 68.98%.To verify the generalization of the improved DlinkNet model in other rural regions within the agricultural landscape, Gaoyi County in Hebei Province is taken as an example. The farmland road extraction task is completed, and an in - depth analysis is performed based on the extraction results. Through research on optimizing the location selection of straw collection and storage stations, the straw yield and theoretical available amount of winter wheat, key parameters in agricultural production assessment, are calculated. Based on the World Cover dataset, this study optimized the site selection of straw collection and storage stations using the K-means clustering algorithm. By defining straw resource distribution density, residential areas, rivers, lakes, and other environmental factors, the optimal locations for village-level straw collection stations were determined. The results demonstrate that the selected sites meet the requirements of environmental protection, straw resource availability, and transportation accessibility, laying the foundation for subsequent research on straw transportation route optimization. This contributes to the establishment of a comprehensive optimization system for straw collection, storage, and transportation.