Porosity is a critical factor affecting the windbreak and sand-fixing performance of upright reed sand fences. However, achieving consistent porosity control is challenging due to variations in reed bundle morphology and manufacturing complexities. This paper presents an intelligent system that integrates deep learning with realtime control to accurately regulate porosity. A lightweight instance segmentation model, Reed-YOLOv8n-SSAS, was developed for machine vision-based monitoring. Compared to the baseline YOLOv8n-seg, it achieved 98.6 % precision and 98.7 % mAP@0.5, while reducing the model size, parameter count, and computational cost by 51.5 %, 55.4 %, and 40.0 %, respectively. A mask relocation algorithm was proposed to estimate the equivalent feed diameter during the transition from dispersed to bundled reeds. This enables real-time adjustment of the forming machine's parameters via a closed-loop control strategy. Field tests demonstrated that the system reliably maintained porosity near the target 45 %. Validation along the Xinjiang He-Ruo Railway showed that mechanically produced fences exhibited high structural integrity (straightness deviation <= 3 degrees) and consistent porosity (44.8 +/- 1.2 %), with significantly higher production efficiency compared to manual methods. This study provides a viable technical pathway for the green, efficient, and intelligent manufacturing of materials for desertification control.