To improve the accuracy of rice pest and disease detection and address issues such as the numerous types of pests and diseases, varying scales, and complex growth environments that lead to false positives and missed detections, while also facilitating model deployment on edge devices, we propose a rice pest and disease detection method based on an improved YOLOv8, called PBM-YOLOv8. Firstly, the Partial Convolution (PConv) is used to design the PCBlock structure, replacing the bottleneck structure in the YOLOv8 feature extraction module to reduce model parameters and improve detection speed. Secondly, to minimize the dilution of semantic information during feature fusion between non-adjacent layers, a balanced feature fusion layer is added to the neck network to reconstruct the feature pyramid into a balanced feature pyramid (BFP). This involves re-extracting features from the fused feature layers and introducing Embedded Gaussian Non-local Attention (EGNA) to eliminate aliasing effects caused by multi-layer fusion, thereby minimizing feature loss. Finally, the loss function is replaced with MPDIOU to address detection box distortion caused by significant sample differences while reducing the computational burden of model training. Experimental results show that the improved PBM-YOLOv8 model achieves superior results on the rice pest and disease dataset, with precision (P) and mean average precision (mAP) increased by 1.3 and 1.1 percentage points, respectively, compared to the original YOLOv8n baseline model. After multi-thread optimization, the detection speed of PBM-YOLOv8 deployed on RK3588 can reach 71.4 FPS, meeting the requirements of practical applications for real-time and accurate detection of rice pests and diseases.