Objective To address the challenges of small-object detection difficulty and insufficient feature representation in Oryza sativa L. (rice) pest identification, this study proposed an improved rice pest detection method, termed DSDC-YOLO, to enhance small-object detection accuracy in complex environments.

Method First, the C3k2 module was replaced with the proposed C3k2-DRB to enhance the modeling of sparse long-range dependencies and spatial features. Second, spatial depth convolution (SPDConv) was employed instead of conventional downsampling to maximally preserve fine-grained pest details. Third, to enable efficient interaction between shallow fine-grained information and deep semantic features, a dynamic range gated fusion (DRGFusion) mechanism was integrated before the detection head. Finally, a compound focal weighted IoU (CFW) loss function was designed to improve classification performance under class imbalance and enhance bounding box localization accuracy.

Result Experiment results demonstrated that DSDC-YOLO achieved precision, recall, and mAP@0.5 of 86.7%, 79.1%, and 86.4%, respectively, on the rice pest dataset, outperforming YOLO11 by 3.2, 0.9, and 2.4 percentage points. In comparisons with mainstream models, including YOLOv5, YOLOv8, and YOLOv10, the proposed method consistently delivered superior overall performance. For small object scenarios, DSDC-YOLO attained an mAP@0.5 of 40.2% with competitive recall, indicating enhanced capability in detecting small-scale targets.

Conclusion The DSDC-YOLO model demonstrates robust detection performance under diverse conditions, including strong illumination and low-light environments. It achieves verifiable improvements in small pest detection and provides a reproducible technical framework and application reference for intelligent rice pest recognition in agriculture.