A Multi-Scale Attention U-Net for Extracting Agricultural Greenhouses from Very High Resolution Imagery

With advancements in remote sensing technology, very high resolution imagery plays a vital role in agricultural monitoring. Specifically, the extraction of agricultural greenhouses from satellite imagery is critical for precision farming and land management. However, existing methods face challenges due to the lack of specialized datasets and the poor adaptability of generic segmentation models to greenhouse-specific features (e.g., irregular shapes and complex backgrounds). To address these issues, this study first constructs a large-scale greenhouse segmentation dataset using China's Jilin-1 satellite imagery (0.75 m spatial resolution), comprising 4,529 annotated images covering diverse terrain environments. Furthermore, we propose a MultiScale Attention U-Net (MAS-U-Net), which is an improved U-Net model integrating multi-scale feature extraction, Channel-Spatial Efficient Channel Attention(CS-ECA), and hierarchical feature fusion. Experiments demonstrate that MAS-U-Net achieves a mean Intersection over Union (mIoU) of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{8 9. 8 2 \%}$</tex> and mean Pixel Accuracy (mPA) of 94.23%, outperforming U-Net, FCN, and DeepLab by <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$4.48 \%, 3.57 \%$</tex>, and 2.02 % in mIoU, respectively. Additionally, the model maintains real-time inference efficiency (0.6 seconds per image). This work establishes a methodological insight for greenhouse mapping and improves the application of high-resolution satellite data in precision agriculture.