A multi-feature fusion based remote sensing inversion method for farmland shelterbelts

Precise segmentation of farmland shelterbelts in high-resolution remote sensing imagery represents a crucial yet challenging task for establishing a quantifiable farmland quality evaluation system. The core difficulties arise from two principal issues: (1) effectively distinguishing cultivated land from shelterbelts with similar textural characteristics while suppressing interference from complex backgrounds such as roads and ditches; and (2) accurately segmenting narrow, elongated, and discontinuously distributed single-row shelterbelts with blurred boundaries. Conventional semantic segmentation methods, primarily designed for large-scale objects in natural scenes, generally underperform when confronted with the distinctive characteristics of remote sensing targets. To overcome these challenges, we propose a novel remote sensing inversion framework based on multi-feature fusion. For the first challenge, we designed a Multi-Feature Fusion Block (MFFB) that utilizes a Spatial Gated Fusion Mechanism (SGFM) to adaptively integrate global contextual features captured by Mamba-like linear attention, local details extracted through convolutional operators, and frequency-domain information obtained via Fast Fourier Transform (FFT), thereby significantly enhancing the model’s capacity to represent and discriminate complex features. To address the second challenge, we introduced a super-resolution preprocessing strategy along with a Multi-Scale Contextual feature Extraction (MSCE) module within an encoder-decoder architecture. The former effectively increases the pixel width of narrow shelterbelts through enhanced image detail reconstruction, while the latter ensures segmentation continuity for elongated features by integrating multi-scale contextual information. Experimental results on our self-constructed farmland shelterbelt dataset demonstrate that our method achieves segmentation accuracies of 96.42% for cultivated land and 82.83% for shelterbelts, outperforming both mainstream general-purpose semantic segmentation models and specialized remote sensing methods, thus validating the effectiveness of the proposed framework for precise farmland shelterbelt extraction.