Abstract: Wheat leaf yellowing is a critical phenotypic trait that directly reflects the nutritional status, physiological stress response, and growth vitality of wheat. Abnormal yellowing leads to a sharp decline in chlorophyll content and photosynthetic efficiency, which not only impairs dry matter accumulation but also significantly reduces grain yield and quality. Accurate and real-time segmentation of wheat leaf yellowing areas, combined with scientific severity grading, constitutes a core technical requirement for intelligent growth monitoring, precise nutritional stress diagnosis, and targeted field management. However, existing research on wheat leaf phenotypic analysis mainly focuses on disease type identification and leaf area measurement, with limited in-depth investigation into the fine-grained segmentation of gradual yellowing regions and the associated severity grading. Moreover, most current segmentation models suffer from large parameter counts, high computational complexity, and poor real-time performance, hindering their deployment on resource-constrained edge devices and restricting on-site detection of wheat leaf yellowing. To address the these challenges, this study proposes a lightweight wheat leaf yellowing segmentation and grading model, termed CTGE-MobileUNet (color-texture guided edge-refined MobileNetV3-UNet). Taking the lightweight MobileNetV3-large as the encoding backbone, the model adopts depthwise separable convolutions and inverted residual structures to reduce the parameter and computational complexity while preserving effective multi-scale feature extraction of wheat leaves. A color-texture guided attention (CTGA) module is constructed and embedded into the shallow, middle, deep, and bottleneck layers of the encoder. By integrating multi-channel information - including texture features, HSV, LAB, and RGB - the CTGA module enhances the model's capability to perceive features of gradual yellowing regions with low contrast and blurred boundaries. In the decoding stage, an edge refinement module (ERM) is incorporated to strengthen the structural representation at the junctions between wheat leaf veins and yellowing regions, effectively mitigating edge blurring and discontinuity in segmentation results. Additionally, sub-pixel convolution is used to replace traditional interpolation and deconvolution operations, improving the model's spatial resolution recovery capability through channel rearrangement and reducing the generation of segmentation artifacts. To evaluate the performance of the proposed model, comparative experiments were conducted against three original models, UNet, PSPNet, and DeepLabV3+. Experimental results show that the CTGE-MobileUNet model achieves fine segmentation of wheat leaf yellowing areas through multi-module collaboration. Compared with UNet, PSPNet, and DeepLabV3+, it improves the mean intersection over union (mIoU), mean pixel accuracy (mPA), and F1-Macro by 2.53～7.11, 2.98～6.89, and 1.90～5.67 percentage points, respectively, indicating excellent fine segmentation performance for wheat leaf yellowing areas. Furthermore, as shown by the loss function and mIoU curves, the CTGE-MobileUNet model converges steadily under the freeze-thaw staged training strategy. The loss function decreases from an initial value of approximately 0.95 to around 0.65, and the validation mIoU eventually stabilizes at approximately 80%, demonstrating that the model possesses good generalization ability and segmentation accuracy, enabling effective and precise segmentation of wheat leaf yellowing areas. In terms of lightweight performance, the parameter count of CTGE-MobileUNet is only 11.19 M, with a prediction time of 9.04 ms. Compared with UNet, PSPNet, and DeepLabV3+, the parameter count is reduced by 13.70, 35.52, and 43.52 M, respectively, achieving a favorable balance between segmentation accuracy and model efficiency. Based on the lesion area ratio method specified in the national standard GB/T15790-2017, experiments on severity grading of wheat leaf yellowing were conducted. The average grading accuracy of the CTGE-MobileUNet model reaches 93.08%, which is 3.74, 10.24, and 2.21 percentage points higher than that of UNet, PSPNet, and DeepLabV3+, respectively, indicating that the model possesses robust stability and discriminative ability in the task of quantitative severity grading. For practical edge deployment, the CTGE-MobileUNet model was deployed on the NVIDIA Jetson Orin Nano edge computing device and optimized with TensorRT acceleration. Deployment results show that without acceleration, the average prediction time on the edge device is 0.17 s with 14.96 G floating-point operations (FLOPs). After TensorRT acceleration, the FLOPs of the model remain stable, while the average prediction time is reduced to 0.11 s, a decrease of 35.29% compared with that before acceleration, which significantly improves the real-time inference performance of the model on edge devices. In conclusion, the edge device deployment and TensorRT acceleration optimization completed in this study validate the practical applicability of the CTGE-MobileUNet model. Meanwhile, the model exhibits notable advantages in lightweight design, gradual yellowing region segmentation, boundary structure restoration, and severity grading, achieving an excellent balance between segmentation accuracy and inference efficiency. This study confirms the feasibility and effectiveness of the proposed model under controlled laboratory conditions, laying a solid technical foundation for future large-scale field identification and intelligent evaluation of wheat leaf yellowing severity.