Prediction of hydraulic blockage at culverts from a single image using deep learning

Abstract Cross-drainage hydraulic structures such as culverts and bridges in urban landscapes are prone to get blocked by the transported debris (e.g., urban, vegetated), which often reduces their capacity triggers flash floods. Unavailability of relevant data from blockage-originated flooding events complex nature accumulation highlighted factors hindering research within blockage management domain. Wollongong City Council (WCC) conduit policy is leading formal guidelines incorporate into design guidelines; however, criticized engineers for its dependence on post-flood visual inspections (i.e., blockage) instead peak floods investigations blockage). Apparently, no quantifiable relationship reported between blockage; therefore, many consider WCC invalid. This paper exploits power Artificial Intelligence (AI), motivated recent success, attempts relate with proposing a deep learning pipeline predict an image culvert. Two experiments performed where conventional end-to-end approaches implemented compared context predicting single image. In experiment one, approach feature extraction using CNN regression ANN) adopted. contrast, two, models E2E_ MobileNet, BlockageNet) trained approach. Dataset Hydraulics-Lab Blockage (HBD), Visual (VHD)) used this were collected laboratory scaled physical culverts. BlockageNet model was best $$R^2$$ R 2 score 0.91 indicated that could be interrelated features at