Using the LADI dataset, along with other open data sets and open source models, to develop deep learning models for detection of flooding and infrasturcture and to compose a tutorial for it.
Sponsored By: Massachusetts Institute of Technology (MIT)
Team Members
Nae-Rong Chang | Medha Deeptimahanti | Kiana Harris | Jianyu Mao | Caleb Pennel | Yiming Ren | | | | | |
Project Poster
Click on any image to enlarge.
Watch the Project Video
Download the Project Summary
Project Summary
Overview
Researchers from MIT’s Lincoln Laboratory began assessing various disaster damages after Hurricane Maria struck Puerto Rico. Human and machine aerial images of various disasters were then collected by the Civil Air Patrol from 2015-2019. This dataset was titled as Low Altitude Disaster Imagery (LADI). In the initial release of the dataset, there were far too many images to sort through manually and the recognition systems were not properly identifying the images. In any large-scale disaster scenario, teams of emergency responders could save significant time and resources by reviewing the conditions prior to arriving.
Objectives
The main objective of this project was to develop a deep learning model utilizing the LADI dataset to classify images and prioritize classification of flooding, debris, and buildings/infrastructure. While experimenting with different deep learning models using PyTorch, another objective was to create tutorials and documentation that clearly described the steps we took to accomplish the main goal of this project. Since we are the first team to develop classifiers for this dataset, our work will aid MIT Beaver Works Summer Institute course for Remote Sensing for Disaster Response. All code and documentation created is publicly available on GitHub for universal accessibility and enables potential end-users to adopt and modify the models.
Approach
We took the following steps to approach our problem:
– Configure AWS environment and load the LADI dataset into an Amazon S3 bucket and Python on a local machine
– Select 2000 random images for training to clean an aggregated response file and filter them based on human generated labels of True/False of flood water images
– Load the dataset generated into a Pytorch framework to iterate through the flood water images to create batches and implement image transforms using DataLoader
– Trained a binary classifier to classify flood water images using CNN, ResNet, DenseNet, MobileNet and AlexNet achieving frequencies between 64% to 79%.
We concluded that ResNet 101 is our best model as it resulted in the highest accuracy of 79% with a model size of 324.8 MB.
Outcomes
– Achieved above a 70% accuracy score for ResNet34, ResNet50, ResNet101, DenseNet161, MobileNetV2, and AlexNet models.
– MobileNet, which is best deployable on hardware due to its small size of 17MB, achieved an accuracy of 73%.
– Thorough and detailed documentation explaining each step taken throughout the project, which allows potential end-users to adopt, modify, and even improve our already existing work.