Semantic Segmentation of Flooding for Flooding Detection
Description/Abstract/Artist Statement
Among all natural disasters, flooding occurs frequently around the world, causing significant harm to humans, livestock and property. The overall goal for this project is to establish a real-time flood detection system that can warn motorists of flooding in relevant areas. Semantic segmentation of water is naturally required to build such a flood detection system. In this work, we present a semantic segmentation model trained to segment water in images and videos of flooding. We first train it on a unique combination of all preexisting semantic segmentation datasets with water and perform manipulations and mappings to these datasets to prepare them for training. Afterwards, we perform a grid search to find the most optimal neural network and corresponding loss function for the prepared dataset. The result is a neural network that can efficiently and accurately detect water in images and videos of flooding. Our model converged to a binary accuracy of 87%, mIOU of 56%, F1 score of 67%, F2 score of 70%, precision of 70%, and recall of 76% on our prepared dataset. These statistics translate well into reality as indicated by the video linked here with blue representing flooding and red representing not flooding: https://www.youtube.com/watch?v=UIAW1r1Zqco&ab_channel=StephenLamczyk
Faculty Advisor/Mentor
Khan Iftekharuddin
College Affiliation
College of Engineering & Technology (Batten)
Presentation Type
Oral Presentation
Disciplines
Artificial Intelligence and Robotics | Vision Science
Session Title
Interdisciplinary Research #1
Location
Zoom
Start Date
3-19-2022 2:15 PM
End Date
3-19-2022 3:15 PM
Semantic Segmentation of Flooding for Flooding Detection
Zoom
Among all natural disasters, flooding occurs frequently around the world, causing significant harm to humans, livestock and property. The overall goal for this project is to establish a real-time flood detection system that can warn motorists of flooding in relevant areas. Semantic segmentation of water is naturally required to build such a flood detection system. In this work, we present a semantic segmentation model trained to segment water in images and videos of flooding. We first train it on a unique combination of all preexisting semantic segmentation datasets with water and perform manipulations and mappings to these datasets to prepare them for training. Afterwards, we perform a grid search to find the most optimal neural network and corresponding loss function for the prepared dataset. The result is a neural network that can efficiently and accurately detect water in images and videos of flooding. Our model converged to a binary accuracy of 87%, mIOU of 56%, F1 score of 67%, F2 score of 70%, precision of 70%, and recall of 76% on our prepared dataset. These statistics translate well into reality as indicated by the video linked here with blue representing flooding and red representing not flooding: https://www.youtube.com/watch?v=UIAW1r1Zqco&ab_channel=StephenLamczyk