Abstract: Farmland shelterbelts, presenting as typical narrow linear features in remote sensing imagery, pose significant challenges for accurate segmentation due to their strong global contextual dependencies and weak local characteristics. To achieve high-precision automated extraction of both cultivated land and shelterbelts, this study proposes a novel semantic segmentation model named MFF-Net. The core of our approach is the design of a Multi-Feature Fusion Block (MFFB), which incorporates a Spatially Gated Fusion Mechanism (SGFM) to adaptively integrate three complementary capabilities: the long-range dependency modeling of a Mamba-like Linear Attention (MLLA) operator, the local detail perception of convolutional networks, and the frequency-domain edge enhancement afforded by Fast Fourier Transform (FFT). This integrative design effectively addresses the feature representation imbalance often encountered in segmenting elongated objects. Furthermore, we introduce a task-oriented super-resolution preprocessing pipeline. This front-end step is designed to reconstruct high-resolution images, thereby enhancing textual details and sharpening boundaries of subtle features like shelterbelts, providing superior input for the subsequent segmentation model. Comprehensive experiments on a self-constructed farmland shelterbelt dataset demonstrate the superior performance of MFF-Net. Our model achieved precision rates of 96.42% for cropland and 82.83% for shelterbelts, with a mean Intersection over Union (mIoU) of 83.45%, surpassing a range of advanced models including RS³Mamba, DC-Swin, DeepLabV3+, and SegMAN. Ablation studies systematically validated the effectiveness of each component within our multi-feature fusion strategy. The introduction of frequency-domain features via FFTFormer contributed to a 2.80% increase in shelterbelt IoU, highlighting its role in enhancing boundary discrimination. The Spatially Gated Fusion Mechanism (SGFM) further boosted the overall mIoU by 1.71%, confirming its efficacy in adaptively balancing the contributions from different feature domains. Compared to a baseline model, the full MFFB design led to a substantial overall improvement of 5.14% in mIoU. The integration of super-resolution preprocessing proved to be a highly effective auxiliary strategy. Using 4x upsampled images as input resulted in a remarkable 6.61% increase in shelterbelt IoU and a 2.47% gain in overall mIoU, effectively mitigating the difficulties associated with segmenting narrow targets by augmenting their pixel-width and edge clarity. To demonstrate practical utility, we established a complete technical pipeline from pixel-level segmentation to vectorization and application. The vectorized results enabled accurate area estimation, with average relative errors of 7.50% for cropland and 6.76% for shelterbelts compared to national survey data. An application analysis of shelterbelt closure degree for individual farmland plots further showcased the method's potential, revealing that 83.71% of the plots met the national standard requirement (≥0.75). In conclusion, the MFF-Net model, through its synergistic fusion of global, local, and frequency-domain features combined with task-specific super-resolution enhancement, effectively tackles the challenges of segmenting narrow linear features in complex agricultural landscapes. This research provides a robust and automated technical pathway for the precise monitoring and dynamic assessment of farmland shelterbelt networks, offering significant value for cultivated land quality evaluation and ecological conservation.