Automated Detection of Hazards
Application of Conventional Neural Networks in Location of Hurricanes
Summary
Ingenii chose to locate Hurricanes as a solution to the Automated Detection of Hazards. Ingenii trained a conventional neural network (CNN) to located and use Transverse Cirrus Bands (TCB) and Ionosphere to look for hurricanes. The neural network coded assuming an input layer of a picture of 200x200 resolution has ~80 million params. The model had achieved an accuracy of 75% on a 90/10 training/test data split into 2000 data points. The model's 75% accuracy is great in relation to paper studies of TCB bands which achieved an accuracy of ~80%. The CNN can further be improved with more computational resources by increasing the number of iterations and epochs, and adjusting the Gaussian blur.
How We Addressed This Challenge
This program was a conventional neural network aimed to be an early warning system of hurricanes and tropical storms. This is important for hurricanes cause billions of dollars worth of damage annually and more tragically and have a heavy death toll. This model achieved an operational accuracy of 75%, while further room for improvement through adjusting the the hyper parameters of the Gaussian blur in the image preprocessing and increasing the epoch and number of iterations. The input layer of the conventional neural network in a function of NASA's Transverse Cirrus Bands Data and CSA's CASSIOPE data in relation to the Ionosphere. Furthermore the GUI uses data from NASA's GIBS for a visual aid.
How We Developed This Project
The team started the process with initial research through a combination of image processing and image recognition. Through some discussion we landed on writing the code in python for the abundance of publicly available resources and the quickly approaching deadline of a couple days. Quickly through some failed initial testing of algorithm solutions artificial intelligence was quickly shown to be the best area to move forward in. Originally the team used popular AI modules such as Sklearn however through extensive testing with was evident that the unoptimized Sklearn would have to be replaced, after from other modules we ended up landing on Tensorflow, and using OpenCV for image preprocessing. Tensorflow and OpenCV documentation was read to get an innate understanding to the processes. Originally the team worked together on google colab servers for the great computational resources and user support. However for development of the GUI the code was moved onto the private systems. The team faced 100 failure to 1 success on creation of the AI, the number of hidden layers, what the input layer should be, the number of epoch, the activation function, the dense layers, the flattening of the input, kernel size, and filters. Finally we had some rudimentary success in the way of getting above a benchmark of 50% accuracy. Overall, the team is proud of the project and a lot smarter than when we commenced.
How We Used Space Agency Data in This Project
CSA's CASSIOPE - Input Layer (Was instrumental in the accuracy of the program)
NASA's TCB - Input Layer (Allowed for the input layers accuracy)
NASA's GIBS - GUI and Output Layer (Visual aid in the GUI, and when data for predicting location of hurricanes is from the GIBS)
Project Demo
LINK TO CODE: https://github.com/stefanZorcic/Hurricane_CNN
https://docs.google.com/document/d/1E0qWJ5jUJouL6BBwoUxNYnWQ0w6naKDnKFlHhquRcFg/edit?usp=sharing
Data & Resources
CSA's CASSIOPE
NASA's TCB
NASA's GIBS