Machine Learning Detection of Collision-Risk Asteroids

Abstract

Asteroids have attracted people's attention from the past to the present. It has a wide place in the beliefs and cultures of ancient civilizations. The sense of discovery and curiosity of human beings causes an increase in their interest in these objects. With the technology coming to a certain level, the detection, diagnosis and materials of asteroids can be found clearly. The route and collision effects of these objects require constant observation. In our study, asteroids that are likely to hit the Earth have been classified using an asteroid data set in Kaggle and the source of which is NASA-JPL. The dataset contains 4687 asteroid data. Pre-processing steps such as filling in missing data, anomaly detection and normalization were applied on the data. Then, with the help of correlation, 19 features were determined from the dataset for dangerous situations. Asteroid classification was made by using Decision Tree with features, Naive Bayes, Logistic Regression, Random Forest, Support Vector Machines, K-Nearest Neighbor, Xgboost and Adaboost machine learning algorithms. With the artificial neural network with different number of neurons and layers, the data were trained and compared with classification algorithms. As a result of the comparison, the highest performance was achieved with the AdaBoost algorithm with 99.80%. Hyperparameter optimization was performed using the grid-search method in all the classification algorithms that were run. Thus, a method that requires continuous observation and enables the processing of large amounts of data in a more efficient way has been proposed.