Abstract: Litchi is one of the highly favorite fruits in China, due to its tender, succulent, and sweet pulp. It is essential to the rapid and accurate identification of the pests and diseases in litchi orchards. However, the natural factors (such as lighting conditions and wind) can significantly interfere with the detection of the litchi pests and diseases in real-world environments of the litchi orchard. Additionally, the advanced detection models are often required for the stable yields, due to the limited computing of edge devices. In this study, an improved lightweight YOLOv8n model was proposed to identify the litchi pests and diseases using the YOLOv8n network framework. Firstly, the neck network of the YOLOv8n architecture was enhanced to implement a lightweight cross-scale feature fusion module (CNN - based Cross - Scale Feature Fusion, CCFM). There was the effective reduction in the number of model parameters, computational complexity, and model size. Meanwhile, a lightweight dynamic upsampler (DySample) was introduced into the neck network, in order to offset the potential accuracy loss that caused by the lightweight design. Secondly, the Receptive-Field Attention Convolution (RFAConv) was integrated into the backbone network of YOLOv8n. Furthermore, the more refined RFAConv with the lightweight was used to extract the features of the litchi pests and diseases in natural environments. There was the relatively small interference of the natural environmental factors on the model detection. Finally, the Inner-MPDIoU (combining Inner - IoU with MPDIoU) loss function was employed to replace the original loss function. The same aspect ratio but different specific sizes were observed in the predicted and ground truth bounding boxes. Experimental results demonstrated that the improved model was achieved a reduction of 33.3%, 30.6%, and 16.0% in the number of the parameters, model size, and computational amount, respectively, compared with the original YOLOv8n benchmark model. Meanwhile, the precision, recall, and mean average precision increased by 2.1 percentage points, 9.1 percentage points, and 4.5 percentage points, respectively. The improved model was deployed on Jetson Nano and Jetson Orin NX development boards, where the TensorRT was used to accelerate the detection. There were the detection speeds of 35.8 frames per second and 67.7 frames per second, respectively, fully meeting the requirements of the real-time detection. The improved model was significantly reduced the missed detections and false positives, more suitable for the precise identification of the litchi pests and diseases in natural environments. Moreover, the improved model with the fewer parameters and higher detection speed was deployed on the fixed camera devices in orchard. The real-time image acquisition was obtained to accurately identify the types, occurrence areas, and damage levels of the pests and diseases in litchi orchards. Dynamic early warnings were also offered for the timely decision-making support. The intelligent control systems of plant protection drones were integrated for the real-time perception of the pest and disease distributions. The "on-demand supply" of the pesticides was realized to precisely plan the flight paths and spraying strategies of the drones. The spraying dosage increased in the high-incidence areas. The lightly affected areas were reduced to avoid the pesticide waste and environmental pollution that caused by the traditional "full - coverage" spraying. The finding can provide the technical support for the efficient and safe production of the litchi industry. A strong reference and practice can also be offered for the intelligent prevention and control of the crops pests and diseases in smart agriculture.