Doctor of Philosophy
SMART Infrastructure Facility
Floods are the most recurrent, widespread and damaging natural disasters, and are ex-pected to become further devastating because of global warming. Blockage of cross-drainage hydraulic structures (e.g., culverts, bridges) by ﬂood-borne debris is an inﬂuen-tial factor which usually results in reducing hydraulic capacity, diverting the ﬂows, dam-aging structures and downstream scouring. Australia is among the countries adversely impacted by blockage issues (e.g., 1998 ﬂoods in Wollongong, 2007 ﬂoods in Newcas-tle). In this context, Wollongong City Council (WCC), under the Australian Rainfall and Runoff (ARR), investigated the impact of blockage on ﬂoods and proposed guidelines to consider blockage in the design process for the ﬁrst time. However, existing WCC guide-lines are based on various assumptions (i.e., visual inspections as representative of hy-draulic behaviour, post-ﬂood blockage as representative of peak ﬂoods, blockage remains constant during the whole ﬂooding event), that are not supported by scientiﬁc research while also being criticised by hydraulic design engineers. This suggests the need to per-form detailed investigations of blockage from both visual and hydraulic perspectives, in order to develop quantiﬁable relationships and incorporate blockage into design guide-lines of hydraulic structures. However, because of the complex nature of blockage as a process and the lack of blockage-related data from actual ﬂoods, conventional numerical modelling-based approaches have not achieved much success.
The research in this thesis applies artiﬁcial intelligence (AI) approaches to assess the blockage at cross-drainage hydraulic structures, motivated by recent success achieved by AI in addressing complex real-world problems (e.g., scour depth estimation and ﬂood inundation monitoring). The research has been carried out in three phases: (a) litera-ture review, (b) hydraulic blockage assessment, and (c) visual blockage assessment. The ﬁrst phase investigates the use of computer vision in the ﬂood management domain and provides context for blockage. The second phase investigates hydraulic blockage using lab scale experiments and the implementation of multiple machine learning approaches on datasets collected from lab experiments (i.e., Hydraulics-Lab Dataset (HD), Visual Hydraulics-Lab Dataset (VHD)). The artiﬁcial neural network (ANN) and end-to-end deep learning approaches reported top performers among the implemented approaches and demonstrated the potential of learning-based approaches in addressing blockage is-sues. The third phase assesses visual blockage at culverts using deep learning classiﬁ-cation, detection and segmentation approaches for two types of visual assessments (i.e., blockage status classiﬁcation, percentage visual blockage estimation). Firstly, a range of existing convolutional neural network (CNN) image classiﬁcation models are imple-mented and compared using visual datasets (i.e., Images of Culvert Openings and Block-age (ICOB), VHD, Synthetic Images of Culverts (SIC)), with the aim to automate the process of manual visual blockage classiﬁcation of culverts. The Neural Architecture Search Network (NASNet) model achieved best classiﬁcation results among those im-plemented. Furthermore, the study identiﬁed background noise and simpliﬁed labelling criteria as two contributing factors in degraded performance of existing CNN models for blockage classiﬁcation. To address the background clutter issue, a detection-classiﬁcation pipeline is proposed and achieved improved visual blockage classiﬁcation performance. The proposed pipeline has been deployed using edge computing hardware for blockage monitoring of actual culverts. The role of synthetic data (i.e., SIC) on the performance of culvert opening detection is also investigated. Secondly, an automated segmentation-classiﬁcation deep learning pipeline is proposed to estimate the percentage of visual blockage at circular culverts to better prioritise culvert maintenance. The AI solutions proposed in this thesis are integrated into a blockage assessment framework, designed to be deployed through edge computing to monitor, record and assess blockage at cross-drainage hydraulic structures.
Iqbal, Umair, Application of Artiﬁcial Intelligence Approaches in the Flood Management Process for Assessing Blockage at Cross-Drainage Hydraulic Structures, Doctor of Philosophy thesis, SMART Infrastructure Facility, University of Wollongong, 2022. https://ro.uow.edu.au/theses1/1340
FoR codes (2008)
080104 Computer Vision, 080106 Image Processing, 080199 Artificial Intelligence and Image Processing not elsewhere classified, 090509 Water Resources Engineering
Unless otherwise indicated, the views expressed in this thesis are those of the author and do not necessarily represent the views of the University of Wollongong.