Space Apps Challenge 2018

Problem Statement

Develop a machine learning algorithm to correlate the data collected by the GMAG and ASI laboratories to :

1. Improve the detection of auroras in severe weather conditions by using GMAG data as a input for probability measures.

2. Improve the forecasting of auroras through the forecasting on the time series of GMAG data.

3. Develop probabilistic method ( Bayesian trees ) to generate the probability for an aurora occurrence given current and past magnetometer readings, as well as data about past aurora observations.

Initial approaches

1. Time Forecasting methods such as ARIMA, SARIMA, SARIMAX, SARFIMA.

2. Various neural network models such as LSTM, GRU, BiLSTM, Auto-encoder decoders etc.

3. Statistical methods for data correlations.

Team background

Single member, Abhishek . : 3rd year undergraduate student in Computer Science-Physics, Research Analyst in Machine Learning and Data Science.

Problems Faced

Very large datasets which could not be stored locally. To reduce time and effort, used subsets of the given datasets in order to build and test the machine learning models.

Long training times for neural network models due to local machine limitations on GPU memory. Overcame by reducing the network sizes to accomodate training times.

Resources Used

Data :

1. CSA THEMIS Datasets : GMAG and ASI measurement for the year 2017-2018

Programming :

Language - Python 3.6

Package Environment - Anaconda Package Distribution

Libraries used - Keras, TensorFlow, Matplotlib, Pandas, Numpy, os.

IDE - Spyder 3.6

Notebooks - Jupyter Notebooks 3.6

Future steps :

Complete the implementation of Bayesian networks for generating probabilistic output for forecasting the probability on an aurora occurring in the next t timestamps given the current GMAG sensor values.