The secondary discrimination detection method for pipeline corrosion in utility tunnels based on graph fusion

Corrosion of pipeline in utility tunnels is a critical issue affecting the operational lifespan of tunnels, often going unnoticed. To achieve pixel-level detection of corrosion regions in utility tunnel pipelines, this study utilizes a detection vehicle equipped with a hyperspectral imaging camera (400-1000 nm) to collect hyperspectral data from actual tunnel pipelines and performs pixel-level labeling on the regions of interest (ROI). Linear Discriminant Analysis (LDA) is used to extract the optimal classifiable projection direction features of hyperspectral bands based on spectral and labeled data. Principal Component Analysis (PCA) is applied to extract 58, 95, and 129 features corresponding to 85%, 90%, and 95% cumulative variance explanation rates, respectively. Additionally, hyperspectral bands are ranked using the OTSU method based on maximum between-class variance, and the top 50, 100, 150, 200, and up to 250 bands are selected. To address the highdimensionality and noise in hyperspectral data, Support Vector Machine (SVM), Random Forest (RF), and Decision Tree (DT) are chosen as base learners. The outputs of these base learners are then weighted and averaged before being input into a Logistic Regression (LR) meta-learner for final classification, yielding better generalization performance compared to individual learners. The corrosion features obtained from LDA, PCA band extraction, and OTSU band selection are input into an SVM-RF-DT ensemble learning model for training. Experimental results show that, compared to the fullband data, the 230 bands selected using OTSU not only improved the detection speed by 17.75% but also enhanced the classification accuracy to 95.48%, providing the most significant improvement in detecting corrosion pixels. This study provides a fast and efficient solution for pixel-level detection of pipeline corrosion in utility tunnels, with significant practical application value. However, the classification accuracy of edge pixels was relatively low. To address this, based on OTSU classification, the classification probabilities of pixels between 40%-60% were extracted, and 16 texture features (contrast, homogeneity, energy, and correlation) were calculated in four directions (0°, 45°, 90°, and 135°). These features were then input into a K-means model for secondary classification. The results showed that the classification accuracy of edge pixels improved from 54.46% using only spectral features to 87.50% after texture feature fusion. This study provides a fast and efficient solution for pixel-level detection of pipeline corrosion in utility tunnels, with significant practical application value.