Speaker
Description
Seepage is a critical concern in earthen dams, often leading to internal erosion, piping, and potential structural failure if not detected at early stages. Conventional monitoring approaches, such as visual inspection and piezometer measurements, are limited by sparse spatial coverage and a lack of real-time capability. This study proposes a robust and automated framework that integrates infrared (IR) thermography with a dual-input convolutional neural network (CNN) to enable accurate, scalable, and near real-time seepage detection.
The dataset used in this study consists of approximately 500 paired optical and infrared images acquired from controlled laboratory-scale seepage experiments under varying hydraulic and environmental conditions. Each observation includes a co-registered pair of RGB optical images capturing surface texture and structural features, and IR thermograms capturing temperature distributions associated with subsurface moisture movement. The dataset is annotated into two classes: 300 seepage and 200 non-seepage samples. To ensure unbiased evaluation and generalization, the dataset was split into 80% training (400 images) and 20% testing (100 images), maintaining class balance. The IR images exhibit strong diurnal thermal variability, where seepage zones appear as low-temperature anomalies during morning due to evaporative cooling, and as high-temperature anomalies during evening due to thermal inertia. This variability necessitated adaptive preprocessing and labeling strategies.
A comprehensive preprocessing pipeline was implemented, including noise filtering, intensity normalization, and spatial smoothing to reduce sensor noise and environmental artifacts. For enhanced feature separability, clustering-based segmentation using K-means (k = 4) was applied to pixel-level thermal data. The optimal clustering configuration was validated using the Elbow Method and Silhouette Score. Automated cluster labeling was performed based on relative thermal statistics, enabling consistent identification of cold and hot seepage zones.
The processed IR images, along with corresponding optical images, were fed into a dual-input CNN architecture comprising parallel convolutional branches for thermal and spatial feature extraction. The extracted features were concatenated and passed through fully connected layers for binary classification. The network incorporated ReLU activation, batch normalization, and dropout regularization to improve convergence and prevent overfitting. The proposed framework achieved an overall accuracy of 91% on the test dataset. Class-wise evaluation showed a precision of 0.90, a recall of 0.95, and an F1-score of 0.93 for seepage detection, indicating high sensitivity and reliability, while the non-seepage class achieved an F1-score of 0.88. The confusion matrix confirmed minimal false negatives, which is critical for safety-critical applications.
Overall, the integration of advanced preprocessing, structured dataset design, and dual-modal deep learning significantly improves the robustness and accuracy of seepage detection. The framework demonstrates strong potential for real-time deployment and can be extended to UAV-based monitoring systems for large-scale dam health assessment.
