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 hydraulic capacity and triggers flash floods. Unavailability of relevant data from blockage-originated flooding events and complex nature of debris accumulation are highlighted factors hindering the research within the blockage management domain. Wollongong City Council (WCC) blockage conduit policy is the leading formal guidelines to incorporate blockage into design guidelines; however, are criticized by the hydraulic engineers for its dependence on the post-flood visual inspections (i.e., visual blockage) instead of peak floods hydraulic investigations (i.e., hydraulic blockage). Apparently, no quantifiable relationship is reported between the visual blockage and hydraulic blockage; therefore, many consider WCC blockage guidelines invalid. This paper exploits the power of Artificial Intelligence (AI), motivated by its recent success, and attempts to relate visual blockage with hydraulic blockage by proposing a deep learning pipeline to predict hydraulic blockage from an image of the culvert. Two experiments are performed where the conventional pipeline and end-to-end learning approaches are implemented and compared in the context of predicting hydraulic blockage from a single image. In experiment one, the conventional deep learning pipeline approach (i.e., feature extraction using CNN and regression using ANN) is adopted. In contrast, in experiment two, end-to-end deep learning models (i.e., E2E_ MobileNet, E2E_ BlockageNet) are trained and compared with the conventional pipeline approach. Dataset (i.e., Hydraulics-Lab Blockage Dataset (HBD), Visual Hydraulics-Lab Dataset (VHD)) used in this research were collected from laboratory experiments performed using scaled physical models of culverts. E2E_ BlockageNet model was reported best in predicting hydraulic blockage with R2 score of 0.91 and indicated that hydraulic blockage could be interrelated with the visual features at the culvert.

Introduction

     Blockage of cross-drainage hydraulic structures such as culverts and bridges is a commonly occurring phenomenon during floods which often results in a reduced hydraulic capacity of the structure, increased damages to property, diversion of flow, downstream scour, failure of the structure, and risk to life [13, 22, 25, 26, 34, 53,54,55]. Few highlighted examples of blockage-originated floods around the world include Newcastle (Australia) floods [25, 61], Barpeta (India) floods [59], Pentre (United Kingdom) floods [15] and Wollongong (Australia) floods [25, 54]. In the context of Australia, many councils and institutions have mentioned blockage as a critical issue (e.g., NSW Floodplain Management Manual [49], Queensland Urban Drainage Manual [35], Australian Rainfall and Runoff (ARR) [10, 26, 50, 62]), however, none comprehensively addressed consideration of blockage into design guidelines. Research in blockage management is hindered by the highly variable nature of blockage formulation and the unavailability of historical floods data to investigate the behavior of blockage [16, 17, 38]. Wollongong City Council (WCC), under the umbrella of ARR, developed a conduit blockage policy for the first time to incorporate the blockage within the design guidelines [36, 62]. The WCC policy suggested that any hydraulic structure with a diagonal length less than 6m is prone to 100% blockage during peak floods.

Conclusion

     Deep learning pipeline and end-to-end deep learning models have been successfully implemented and compared by performing two experiments in the context of predicting the hydraulic blockage from a single image of the culvert. Experiment one implemented a conventional deep learning pipeline using CNN and ANN to extract the visual features and predict the hydraulic blockage, respectively. MobileNet CNN model with two-layer ANN (i.e., ANN1) was reported best with R2 score of 0.69. Regression performance was observed to be degraded with the increase in the number of extracted visual features, which may be attributed to the presence of increased number of irrelevant and uncorrelated features. Experiment two implemented end-to-end deep learning models to achieve the functionality of the conventional deep learning pipeline and compared the results. From the results of experiment two, the end-to-end learning approach was reported to outperform the conventional pipeline by a significant margin (i.e., R2 of 0.91 for E2E_ BlockageNet in comparison to 0.69 for the conventional pipeline). Improved performance of end-to-end models may be attributed to their capability of self-optimizing the internal components of the network. A positive R2 score for all cases validated the hypothesis of the existence of a relation between visual features of the culvert and corresponding hydraulic blockage. The performance of proposed models is expected to be degraded significantly for the cases where the image contains a background with a similar visual appearance to the debris material blocking the culvert. The development of data pre-processing techniques to mitigate the visual variations caused by other factors (e.g., lighting, debris type, background, weather) is a potential future research direction. Furthermore, deployment of the proposed approach using the AIoT infrastructure for real-world culvert sites is also planned in the near future.