A Prior Knowledge Guided Semi-Supervised Deep Learning Method for Improving Buried Pipe Detection on GPR Data

• Published in: IEEE transactions on geoscience and remote sensing Vol. 62; p. 1
• Main Authors: Ma, Yongjian, Song, Xianmin, Li, Zhihui, Li, Haitao, Qu, Zhaowei
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.01.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Ablation; Annotations; Buried pipe detection; Buried pipes; Data models; Deep learning; Effectiveness; Feature extraction; Ground penetrating radar; Ground Penetrating Radar (GPR); Labels; Location awareness; Machine learning; Object detection; prior knowledge; Quality control; Radar; Radar detection; Semi-supervised learning; Semisupervised learning; structure feature; Training; Unstructured data
• ISSN: 0196-2892; 1558-0644
• DOI: 10.1109/TGRS.2024.3421526

Using a deep learning method to detect buried pipes on Ground Penetrating Radar (GPR) data is popular for its excellent performance potential. But this needs a large amount of high-quality training data which leads to time-consuming and labor-intensive data annotation work. Semi-supervised learning method provides a solution for the situation of small sample size. However, the original semi-supervised learning methods lack control over the quality of pseudo labels during model training. To address the problem, a new prior knowledge-guided semi-supervised deep learning method is proposed to improve the model performance under the small sample. In the method, a prior structure feature is constructed to control the quality of pseudo label during semi-supervised learning. The structure response index is designed to segment the structure subject out for eliminating disturbance from unstructured information. Then the prior structure feature is represented using histograms of oriented gradients and Fourier descriptors to describe the structure's edge-orientation and shape. Based on the prior structure feature, LightGBM is employed as a discriminator to screen out the high-quality pseudo labels for model training. In the experiments, comparison experiments with some state-of-the-art supervised learning methods and semi-supervised learning methods were carried out to validate the method's performance. The ablation studies on labeled dataset size, pseudo label confidence threshold, and prior structure feature were conducted to validate the effectiveness of the prior structural feature in controlling the pseudo label's quality. The results show the prior structure feature is effective, and the proposed method outperforms the other methods.