Date of Award

Summer 2021

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computational Modeling & Simulation Engineering

Program/Concentration

Modeling and Simulation

Committee Director

Mecit Cetin

Committee Member

Khan M. Iftekharuddin

Committee Member

Michel A. Audette

Committee Member

Zhanping Liu

Abstract

Recent research and statistics show that the frequency of flooding in the world has been increasing and impacting flood-prone communities severely. This natural disaster causes significant damages to human life and properties, inundates roads, overwhelms drainage systems, and disrupts essential services and economic activities. The focus of this dissertation is to use machine learning methods to automatically detect floodwater in images from ground level in support of the frequently impacted communities. The ground level images can be retrieved from multiple sources, including the ones that are taken by mobile phone cameras as communities record the state of their flooded streets. The model developed in this research processes these images in multiple levels. The first detection model investigates the presence of flood in images by developing and comparing image classifiers with various feature extractors. Local Binary Patterns (LBP), Histogram of Oriented Gradients (HOG), and pretrained convolutional neural networks are used as feature extractors. Then, decision trees, logistic regression, and K-Nearest Neighbors (K-NN) models are trained and tested for making predictions on floodwater presence in the image. Once the model detects flood in an image, it moves to the second layer to detect the presence of floodwater at a pixel level in each image. This pixel-level identification is achieved by semantic segmentation by using a super-pixel based prediction method and Fully Convolutional Neural Networks (FCNs). First, SLIC super-pixel method is used to create the super-pixels, then the same types of classifiers as the initial classification method are trained to predict the class of each super-pixel. Later, the FCN is trained end-to-end without any additional classifiers. Once these processes are done, images are segmented into regions of floodwater at pixel level. In both of the classification and semantic segmentation tasks, deep learning-based methods showed the best results. Once the model receives the confirmation of flood detection in image and pixel layers, it moves to the final task of finding the floodwater depth in images. This third and final layer of the model is critical as it can help officials deduce the severity of the flood at a given area. In order to detect the depth of the water and the severity of the flooding, the model processes the cars on streets that are in water and calculates the percentage of tires that are under water. This calculation is achieved with a mixture of deep learning and classical computer vision techniques. There are four main processes in this task: (i)-Semantic segmentation of the image into pixels that belong to background, floodwater, and wheels of vehicles. The segmentation is done by multiple FCN models that are trained with various base models. (ii)-Object detection models for detecting tires. The tires are identified by a You Only Look Once (YOLO) object detector. (iii)- Improvement of initial segmentation results. A U-Net like semantic segmentation network is proposed. It uses the tire patches from the object detector and the corresponding initial segmentation results, and it learns to fix the errors of the initial segmentation results. (iv)-Calculation of water depth as a ratio of the tire wheel under the water. This final task uses the improved segmentation results to identify the ellipses that correspond to the wheel parts of vehicles and utilizes two approaches listed below as part of a hybrid method: (i)-Using the improved segmentation results as they return the pixels belonging to the wheels. Boundaries of the wheels are found from this and used. (ii)-Finding arcs that belong to elliptical objects by applying a series of image processing methods. This method connects the arcs found to build larger structures such as two-piece (half ellipse), three-piece or four-piece (full) ellipses. Once the ellipse boundary is calculated using both methods, the ratio of the ellipse under floodwater can be calculated. This novel multi-model system allows us to attribute potential prediction errors to the different parts of the model such as semantic segmentation of the image or the calculation of the elliptical boundary. To verify the applicability of the proposed methods and to train the models, extensive hand-labeled datasets were created as part of this dissertation. The initial images were collected from the web, then the datasets were enriched by images created from virtual environments, simulations of neighborhoods under flood, using the Unity software.

In conclusion, the proposed methods in this dissertation, as validated on the labeled datasets, can successfully classify images as a flood scene, semantically segment the regions of flood, and predict the depth of water to indicate severity

DOI

10.25777/sqnd-rm87

ISBN

9798460436408

Share

COinS