<p>Rice pests and diseases pose a significant challenge to global food security, which is exacerbated by the increasing frequency of climate change and extreme weather events. Conventional manual techniques for the identification of rice pests and diseases are characterized by inefficiency and limited precision. Additionally, they are heavily dependent on pre-existing expertise, rendering them inadequate for application in large-scale detection contexts. In addressing these challenges, this study presents HSFPN-Det, a novel lightweight model developed for the detection of rice pests and diseases. The proposed model integrates a High-level Selective Feature Pyramid Network (HSFPN) with a deformable self-attention mechanism and a Hybrid Attention Transformer, with the objective of improving both detection effectiveness and accuracy. Furthermore, by employing an enhanced network architecture, the neck network is optimized to achieve better multi-scale feature fusion. The deformable self-attention module has been developed to dynamically concentrate on prominent spatial features. The model is evaluated on an augmented rice pest dataset containing 11,292 images across four rice disease categories. In comparative analyses with cutting-edge models, including YOLOv11, YOLOv9, SSD, and EfficientDet-D1, it consistently yields superior performance outcomes. Notably, its compact model size of 3.97 MB is approximately 49.87% smaller than previous models. Additionally, further evaluations conducted on real-world images confirm its superior detection accuracy in diverse scenarios. HSFPN-Det achieves higher accuracy while reducing model complexity, offering a practical and deployable solution for smart agriculture. Its ability to run effectively on low-resource devices marks a crucial contribution to real-time intelligent pest management in precision agriculture.</p>