Multi-scale feature fusion-based semantic segmentation network for agricultural remote sensing images

With the widespread application of high-resolution low-altitude remote sensing technology in agricultural monitoring, fine semantic segmentation of crop plots has become a hot research topic. However, due to the high similarity of land features in large-scale agricultural scenes and the presence of complex situations such as blurred land boundaries, achieving accurate semantic segmentation still faces significant challenges. In response to the above issues, this study proposes a hybrid architecture that combines Convolutional Neural Networks and Transformers, aiming to improve the segmentation accuracy of crop plots in complex scenes, especially in handling areas with fuzzy boundaries and similar features. This method innovatively constructs a global local attention mechanism (GPM-Attention), which generates adaptive attention regions through multi-scale convolution operations, significantly enhancing the model's ability to capture global contextual information. This mechanism not only effectively improves the overall segmentation performance, but also significantly reduces computational redundancy and model complexity by optimizing the computation path. In addition, this study constructed a lightweight edge enhancement module (EEI) as an encoder, which not only expands the local receptive field but also enhances the recognition ability of fine-grained features, effectively solving the problem of crop plot edge blurring. To further optimize the feature fusion effect, this study designed a Feature Adaptive Fusion Module (FAM), which efficiently integrates the multi-level features generated by CNN and Transformer encoders, significantly reducing the semantic information loss of small target features. The experimental results demonstrate that the proposed method achieves a significant performance improvement on the publicly available barley remote sensing dataset, attaining a mean Intersection over Union (mIoU) of 80.39%, which represents an 11.33% increase over state-of-the-art approaches. In addition, the method achieves a 14.2% improvement in F1-score, further confirming its effectiveness. Compared to existing techniques, this study presents a more favorable trade-off among segmentation accuracy, computational efficiency, and model complexity, thereby offering reliable technical support for the practical deployment of low-altitude remote sensing imagery in agricultural monitoring applications.