Li Xiaoyu, Wang AChuan
College of Computer and Control Engineering, Northeast Forestry University, Haerbin, 150040, Heilongjiang, China.
Sci Rep. 2025 Jan 2;15(1):401. doi: 10.1038/s41598-024-84464-3.
Unmanned aerial vehicle (UAV) remote sensing has revolutionized forest pest monitoring and early warning systems. However, the susceptibility of UAV-based object detection models to adversarial attacks raises concerns about their reliability and robustness in real-world deployments. To address this challenge, we propose SC-RTDETR, a novel framework for secure and robust object detection in forest pest monitoring using UAV imagery. SC-RTDETR integrates a soft-thresholding adaptive filtering module and a cascaded group attention mechanism into the Real-time Detection Transformer (RTDETR) architecture, significantly enhancing its resilience against adversarial perturbations. Extensive experiments on a real-world pine wilt disease dataset demonstrate the superior performance of SC-RTDETR, with an improvement of 7.1% in mean Average Precision (mAP) and 6.5% in F1-score under strong adversarial attack conditions compared to state-of-the-art methods. The ablation studies and visualizations provide insights into the effectiveness of the proposed components, validating their contributions to the overall robustness and performance of SC-RTDETR. Our framework offers a promising solution for accurate and reliable forest pest monitoring in non-secure environments.
无人机遥感技术彻底改变了森林病虫害监测和预警系统。然而，基于无人机的目标检测模型容易受到对抗性攻击，这引发了人们对其在实际部署中的可靠性和鲁棒性的担忧。为应对这一挑战，我们提出了SC-RTDETR，这是一种利用无人机图像在森林病虫害监测中进行安全、鲁棒目标检测的新型框架。SC-RTDETR将软阈值自适应滤波模块和级联组注意力机制集成到实时检测变压器（RTDETR）架构中，显著增强了其对对抗性扰动的恢复能力。在真实世界的松树萎蔫病数据集上进行的大量实验表明，与现有方法相比，SC-RTDETR在强对抗攻击条件下平均精度均值（mAP）提高了7.1%，F1分数提高了6.5%，性能优越。消融研究和可视化提供了对所提组件有效性的见解，验证了它们对SC-RTDETR整体鲁棒性和性能的贡献。我们的框架为在非安全环境中进行准确可靠的森林病虫害监测提供了一个有前景的解决方案。
