In order to address the challenges of manual identification of pests in images collected by light traps， as well as the low reliability and poor accuracy of statistical results， this study proposes an improved Cascade R-CNN algorithm for field pest detection. The algorithm is based on the Cascade R-CNN framework and uses ResNeSt-50 as the backbone network， incorporating cross-channel attention mechanisms to obtain feature maps more conducive to pest detection. A unifying object detection head with attentions （DyHead） is used， incorporating scale awareness， spatial position awareness， and task awareness to improve the performance of the detection head. Additionally， the simple copy-paste （SCP） method is employed for data augmentation to enhance the model’s detection capabilities in complex scenarios. A total of 1 500 images of 20 pest categories were collected， and a monitoring lamp field pest dataset compliant with the microsoft common objects in context （MS COCO 2017） format was created. The results show that the F₁-score of the proposed method reaches 86.2%. When the intersection over union （IoU） is set to 0.5， the F₁-score increases by 2.8， 5.8， and 8.2 percentages compared to the classic Cascade R-CNN， Faster R-CNN， and YOLOv4， respectively. The results shows that the proposed method meets the requirements of discriminative ability and real-time performance for monitoring lamp pest detection tasks， achieving high-precision automatic identification and counting of pests， and can be directly applied to field pest detection.