Abstract: Accurate segmentation is often required for the early diagnosis and precise prevention of the crop diseases and pests. However, the existing segmentation of the pests and diseases has been limited to the complex background, fuzzy boundary, and large computation. In this study, a lightweight network with YOLO11-seg was proposed for the pest and disease segmentation. Firstly, the original C3k2 module was replaced with C3k2_FAC in the feature extraction stage, in order to reduce the computational cost for the channel feature response; Secondly, the Focaler-Shape IoU loss function was designed to improve the segmentation accuracy for the complex shapes and edge regions; Thirdly, the LAMP channel pruning was used to further compress the model size and computation; Finally, the channel-level knowledge distillation was utilized for the accuracy recovery, in order to improve the segmentation performance of the model. The dataset consisted of two samples, namely diseases and pests. Among them, the disease samples were collected from the PlantSeg dataset, from which 10 types of disease examples of 5 cash crops, such as corn, wheat, and soybean, were selected as research objects. The pest samples were mainly collected from the network and on-site collection, and then 1 600 pest images were captured. The original dataset of the pests and diseases was expanded to 5 863 using multiple image enhancement techniques, including affine changes, in order to avoid the uneven distribution of data samples and too few samples. After that, the dataset was randomly divided into 4690 training sets, 586 verification sets, and 587 test sets, according to the ratio of 8:1:1. A series of experiments were was conducted using the RTX3090 computing platform with training batches of 32 and 150 iterations on the self-constructed pest and disease dataset. The experimental results show that the optimal performance was achieved after replacing C3k2 in the backbone network with C3k2_FAC. The model parameters and computational load were reduced by 0.7M and 1.3G, respectively, while the mAP increased by 0.4 percentage points. Therefore, the C3k2_FAC module was improved the feature extraction for the high computational efficiency. The Focaler-Shape IoU loss function was introduced, with the mAP₅₀ and mAP_50-95 of 82.7% and 55.0%, respectively. Furthermore, the model size was reduced to 6.0MB, after the LAMP channel pruning was applied with two improvements, significantly decreasing the number of parameters and computational load, although the mAP dropped by 0.5 percentage points. The mAP₅₀ and mAP_50-95 increased by 0.6 and 0.5 percentage points, respectively, after knowledge distillation, in order to balance the model parameters and computational load. The mAP₅₀ and mAP_50-95 of the improved model were 82.7% and 55.7%, respectively, at the mask level, compared with the original model. The size of the model was only 6.3MB, while the number of parameters and the computational volume were reduced by 70.3 and 35.9 percentage points, respectively. There were the outstanding advantages in the segmentation accuracy and lightweight performance, compared with the rest advanced algorithms, such as the Mask R-CNN, SOLOv2, YOLACT, YOLOv8s-seg, and YOLOv9-C. Nine images of the crop disease and pest were selected from the test set. For instance, the segmentation and visual display demonstrated the better recognition on of the pest and disease targets. Leaf diseases were identified to successfully segment on the contours of the lesions, whether they were the single large lesions or multiple small ones. Additionally, the specific shapes and locations of small pests can be expected to accurately distinguish for the high accuracy even amidst relatively cluttered backgrounds. The high accuracy with the lightweight was achieved in the segmentation and identification of the pests. This finding can also provide the technical support to intelligently monitor the agricultural pests and diseases.