CLASSIFICATION OF BOLIDES AND METEORS IN DOPPLER RADAR WEATHER DATA USING UNSUPERVISED MACHINE LEARNING

Abstract

This thesis presents a method for detecting outlier meteors and bolides within Doppler radar data using unsupervised machine learning. Principal Component Analysis (PCA), k-means Clustering, and t-Distributed Statistical Neighbor Embedding (t-SNE) algorithms are introduced as existing methods for outlier detection. A combined PCA and t-SNE method that uses a Nearest Neighbor Density Pruning method for dataset size reduction is also described. These methods are implemented to classify unlabeled radar data from four radar data sites from two bolide events: the KFWS radar for the Ash Creek bolide and the KDAX, KRGX, and KBBX radars for the Sutter’s Mill bolide. The combined PCA + t-SNE method gives an accuracy rate of 99.7% and can classify the data in less than 8 minutes for a 121,000 return sized dataset. However, the classifier’s recall and precision rates remained low due to difficulties in correctly classifying true positive bolides. Some ideas for improving algorithm accuracy, speed, and related follow-on applications are proposed. Overall, the algorithm presented in this research is a viable method to help NASA scientists with bolide detection and meteorite recovery.