The objective of this project is to build an AI that detects the aftermath of a given natural disaster through scene-based images.
Sponsored By: Lockheed Martin Corporation 2
Team Members
Collin Kovacs | Alexander Kim | Connor Kilcoyne | Rebecca Molishus | Dhruv Jaiswal | Tzu-Chieh Huang | Ruofan Yu | | | | |
Project Poster
Click on any image to enlarge.
Watch the Project Video
Download the Project Summary
Project Summary
Overview
Our group was assigned to find a solution that involved creating a machine learning (ML) and artificial intelligence (AI) model to help Search and Rescue Missions by increasing the percentage of successful missions while lowering response time. Our model must be able to accurately predict if an image through aerial and satellite imagery is recognized as a natural disaster or not. The model is able to process images it has not seen before and make predictions based on image densities and other factors.
Objectives
– The user interface shall display each individual image and the entire geospatial world.
– The AI/ML process should search each image and automatically detect if a natural disaster is evident.
– The UI will display the accuracy of the natural disaster recognition.
– The geospatial world should also contain images without natural disasters, people and buildings to represent areas where nothing interesting was found.
Approach
– Design a game board that can show up to 9 images at a time of one of the following disasters: tornadoes, earthquakes, floods, hurricanes and forest fires.
– Let the AI detect whether the images have the features of a selected disaster.
– Gather disasters in two types: natural disaster and “normal” photos with disaster-like characteristics (i.e. an image featuring a bonfire having smoke and flames like in a wildfire image).
– By gathering these types of images, we can let models clearly learn how to accurately detect the specific disaster.
– We designed different kinds of models to predict each disaster.
– After finishing the build of the model, we input new test images that did not exist in the initial data into the models to output data to the game board.
Outcomes
– The sponsor will have the ability to identify natural disasters accurately.
– Since the program is running locally, Lockheed Martin will not have to spend any money on additional resources.
– The process of identifying natural disasters by processing the images within 2 minutes.
– This project uses a novel approach in image scene-recognition classification through a calculation of pixel densities.