CoPix-Row: Structure-aware synthetic data generation for semantic segmentation in precision agriculture

초록

Pixel-level annotation for satellite-imagery-based semantic segmentation is costly, especially for thin and structured targets such as one-pixel crop-row centerlines. This paper proposes convolutional conditional pixel-level row generation, CoPix-Row, a structure-aware synthetic data pipeline that preserves row topology while producing diverse top-view training tiles. A ControlNet-based diffusion generator is conditioned on centerline masks and guided by a layout alignment loss computed with a fixed Row-UNet so that the generated appearance remains faithful to the input geometry. An optional pix2pix-style refiner further adapts texture and noise statistics toward real satellite imagery. A Row-UNet segmenter trained with mixed real and synthetic tiles is evaluated on 50 held-out real test tiles. CoPix-Row achieves an F1-score of 0.521 +/- 0.086 and an IoU of 0.357 +/- 0.077, improving over real-only training by 0.041 in F1-score and 0.036 in IoU, and outperforming the strongest naive mixing baseline based on real and ControlNet-generated data. Additional DeepLabV3+ experiments show that the same trend is preserved across segmentation backbones. Rather than acting as a standalone global path planner, the predicted masks provide planner-consumable structural priors that are qualitatively compatible with existing downstream navigation processes. These results suggest that structure-aware synthesis can reduce labeling burden and improve the reliability of satellite-imagery-based agricultural navigation support under limited real-data conditions.

키워드