Abstract: Wheat leaf yellowing is one of the most serious phenotypic traits to directly reflect the nutritional status, physiological stress response, and growth vitality of the plant. Abnormal yellowing can lead to a sharp decline in chlorophyll content and photosynthetic efficiency, where dry matter accumulation can significantly reduce grain yield and quality. It is often required for accurate and real-time segmentation of wheat leaf yellowing areas. Severity grading can also combine with intelligent growth monitoring, precise nutritional stress diagnosis, and targeted field practices. Existing research on wheat leaf phenotypic analysis has focused on disease type identification and leaf area measurement. However, it is still lacking in the fine-grained segmentation of gradual yellowing regions and the severity grading. Moreover, most current segmentation models suffer from large parameter counts, high computational complexity, and low real-time performance, thus hindering their deployment on resource-constrained edge devices and on-site detection of wheat leaf yellowing. In this study, a lightweight model was proposed for wheat leaf yellowing segmentation and grading, termed CTGE-MobileUNet (color-texture guided edge-refined MobileNetV3-UNet). The lightweight MobileNetV3-large was taken as the encoding backbone. Depthwise separable convolutions were adopted to invert residual structures for fewer parameters and computational complexity, while preserving effective multi-scale feature extraction of wheat leaves. A color-texture guided attention (CTGA) module was constructed and then embedded into the shallow, middle, deep, and bottleneck layers of the encoder. Multi-channel information was integrated, including texture features, HSV, LAB, and RGB. The CTGA module was used to perceive features of gradual yellowing regions with low contrast and blurred boundaries. In the decoding stage, an edge refinement module (ERM) was incorporated to strengthen the structural representation at the junctions between wheat leaf veins and yellowing regions. Edge blurring and discontinuity were effectively mitigated after segmentation. Additionally, sub-pixel convolution was used to replace conventional interpolation and deconvolution. The spatial resolution recovery was improved for fewer segmentation artifacts via channel rearrangement. Comparative experiments were conducted to evaluate the performance of the improved model. Three original models were selected, such as UNet, PSPNet, and DeepLabV3+. Experimental results show that the CTGE-MobileUNet model achieved fine segmentation of wheat leaf yellowing areas after multi-module collaboration. Compared with UNet, PSPNet, and DeepLabV3+, the mean intersection over union (mIoU), mean pixel accuracy (mPA), and F1-Macro were improved by 2.53~7.11, 2.98~6.89, and 1.90~5.67 percentage points, respectively, indicating excellent performance of fine segmentation for the wheat leaf yellowing areas. Furthermore, the CTGE-MobileUNet model converged steadily under the staged freeze-thaw training, as shown by the loss function and mIoU curves. The loss function decreased from an initial value of 0.95 to 0.65. The validation mIoU eventually stabilized at 80%. Better generalization and segmentation accuracy were obtained for the effective and precise segmentation of wheat leaf yellowing areas. In terms of lightweight performance, the parameter number of CTGE-MobileUNet was only 11.19 M, with a prediction time of 9.04 ms. Compared with UNet, PSPNet, and DeepLabV3+, the parameter count was reduced by 13.70, 35.52, and 43.52 M, respectively, indicating a favorable balance between segmentation accuracy and model efficiency. A series of experiments was conducted on the severity grading of wheat leaf yellowing, according to the lesion area ratio specified in the national standard GB/T15790-2017. The average grading accuracy of the CTGE-MobileUNet model reached 93.08%, which was 3.74, 10.24, and 2.21 percentage points higher than that of UNet, PSPNet, and DeepLabV3+, respectively, indicating the robust stability and discriminative ability in the task of severity grading. In practical edge deployment, the CTGE-MobileUNet model was deployed on the NVIDIA Jetson Orin Nano edge computing device and then optimized with TensorRT acceleration. Deployment results show that the average prediction time on the edge device was 0.17 s with 14.96 G floating-point operations (FLOPs) without acceleration. The FLOPs of the model remained stable after TensorRT acceleration. While the average prediction time was reduced to 0.11 s, with a decrease of 35.29%, compared with that before acceleration. which significantly improved the real-time inference performance of the model on edge devices. In conclusion, the edge device deployment and TensorRT acceleration validated the practical applicability of the CTGE-MobileUNet model. Meanwhile, there was an excellent balance between segmentation accuracy and inference efficiency during yellowing region segmentation, boundary structure restoration, and severity grading. The feasibility and effectiveness of the improved lightweight model were also verified under controlled laboratory. The finding can provide a solid technical foundation for large-scale field identification and intelligent evaluation of wheat leaf yellowing severity.