Root morphology is critical for plant adaptation to environmental stress. While deep learning segmentation methods have improved root analysis, they struggle with low-contrast or thin structures in low-resolution images. This paper proposes SRSegFormer, a joint transformer-based framework integrating super-resolution and segmentation via cross-modulation blocks for bidirectional feature refinement. Evaluated on the ChronoRoot dataset, SRSegFormer achieves a Dice of 0.951 and precision of 0.955, outperforming existing baselines and robustly delineating fine and low-contrast root structures. These results confirm the advantage of perceptually guided super-resolution within a transformer segmentation pipeline for high-throughput root phenotyping.