Convolutional neural-network-based automatic dam-surface seepage defect identification from thermograms collected from UAV-mounted thermal imaging camera

Seepage inspections are vitally important for delineating damage zones and ensuring the long-term safe operation of dams. However, vegetation, complex backgrounds, and low signal-to-noise ratio and poor thermogram resolution impose significant adverse effects on automated results. In this study, a novel convolutional neural network is proposed to automatically identify dam-surface seepage from thermograms collected by an unmanned aerial vehicle carrying a thermal imaging camera. An auxiliary input branch with two specially designed modules (i.e., a region-of-key-temperature fusion unit and a convolutional block attention module) are added to a U-Net frame to reduce the false-alarm rate caused by seepage-like background interference on the dams and accurately identify seepage profiles with clear boundaries from the low-resolution thermograms. The method is utilized on actual dams, and experimental results confirm its superiority, even with interference. The Dice coefficient score and intersection-over-union metrics of the proposed network are 87.58 and 78.12 %, respectively: an increase of 3.67 and 4.80% over U-Net and 1.18 and 1.43 % over Mobile DeepLabv3. Under the same computing environment and dataset, the training time of the proposed network in this paper is 8 min and 57 s, U-Net is 3 min and 13 s, Mobile DeepLabv3 is 12 min and 44 s. This study provides a promising and cost-effective automatic dam-surface seepage inspection method.

Automated summary

In this study, a novel convolutional neural network is proposed to automatically identify dam-surface seepage from low-resolution thermograms. The method achieves superior results by reducing false alarms caused by background interference and accurately identifying seepage profiles. Experimental results confirm the effectiveness of the proposed network.