Yunjie Zhang, Dong Ren, Bangqing Chen, and Jian Gu
Object detection, PWD, NAS, feature fusion
Monitoring pine wilt disease(PWD) on remote sensing images is of great significance to the economy and environment. However, there are many problems in this process. When we obtain the images by unmanned aerial vehicle(UAV), because of the changes in mountain height, the diseased tree in the valleys are relatively small in the figure, and the net is easy to ignore the learning of the features of these small diseased trees, resulting in missed detection and unable to be applied in practise. Based on the above problems, this paper proposes a two-stage detection network for PWD in autumn and winter. Specifically, we use the ENF module to fuse the low-level feature maps several times and then use the neural architecture search(NAS)technique to automatically search for the most suitable feature extraction network to better learn the features of the target disease tree. To verify the effectiveness of the method, we conducted ablation experiments and comparative experiments on UAV orthophotos taken near the city of Yichang. Compared to the baseline model, our method improves the mAP and Recall of PWD detection by 5% and 2%, respectively, achieving a 5.4%–6.4% improvement in mAP and 4.6%–17.6% improvement in Recall compared to other models. Experiments have shown that our proposed method can solve the problem of missing PWD in the autumn and winter.
Important Links:
Go Back
