Semantic–Texture Complementation and Prediction-Guided SAM Fusion for Plastic Mulch Segmentation in GF-7 Imagery

Accurate mapping of transparent and white agricultural plastic mulch from very-high-resolution imagery is difficult because bright roofs, roads, and bare soil resemble mulch, while narrow strips and damaged film fragment boundaries. We propose a semantic–texture complementation module (STCM) that adds a partially adapted MobileNetV2 texture encoder and attention-plus-concatenation fusion to PSPNet, DeepLabV3+, and U-Net, and ask whether one jointly selected configuration transfers across the three backbones. Using 1582 georeferenced GF-7 patches partitioned into 1 km spatial blocks with a 100 m exclusion buffer, STCM raised mean foreground F1 by 1.99, 2.30, and 0.99 percentage points for PSPNet, DeepLabV3+, and U-Net, respectively, and improved two-pixel boundary F1 for all three carriers. A capacity-matched scratch-MobileNetV2 control produced lower validation F1 than the ImageNet-initialized condition, associating the gain with pretrained texture representation rather than branch size alone. With STCM–PSPNet as the upstream model, connected components of its prediction generated box and centroid prompts for frozen SAM ViT-H, and a learned FusionNet-S head combined the domain probability with the SAM mask. In pooled summaries, FusionNet-S kept region F1 close to the no-SAM refinement head, improved two-pixel boundary F1 by 2.97 percentage points, and reduced HD95 by 1.19 pixels. A matched randomized-location prompt control retained similar region F1 but did not reproduce the boundary-F1 improvement. Together, the results show that one locked STCM configuration can transfer across the tested CNN carriers and that prediction-guided SAM fusion can strengthen boundary recovery within the evaluated single-date GF-7 domain.