A lightweight attention-enhanced framework for crop pest and disease detection in desert greenhouses

Desert greenhouse environments present challenging visual conditions for crop pest and disease detection because of strong illumination, reflective backgrounds, and frequent leaf occlusion. To address these issues, we propose YOLOv11s-DesertAttn, a lightweight detection framework that integrates ADown, SimAM, and LSKAttention to improve feature preservation, background suppression, and high-level semantic representation. On the test set, the proposed model improves mAP@0.5-0.95 from 0.826 to 0.860 while reducing parameters from 9.43 M to 8.14 M and increasing inference speed from 96 FPS to 116 FPS. Additional evaluation on an independent real-world dataset collected from commercial greenhouses in Hotan, Xinjiang further demonstrates the model's robustness under varying illumination and occlusion conditions, indicating its potential for real-time deployment in intelligent greenhouse monitoring systems.