Electronic Theses and Dissertation

Analisis risiko asteroid telah menjadi perhatian utama dalam bidang astronomi, seiring meningkatnya jumlah data yang dihasilkan oleh teleskop dan satelit. Dalam penelitian ini, telah diimplementasikan algoritma Light Gradient Boosting Machine (LightGBM) untuk mengklasifikasikan asteroid berdasarkan risiko potensial terhadap Bumi dengan menggunakan fitur-fitur multidimensi. Data yang digunakan bersumber dari NASA JPL Small Body Database Query. Penelitian ini bertujuan untuk: (1) menganalisis kinerja LightGBM dalam mengidentifikasi dan memprediksi asteroid berbahaya, (2) memperoleh konfigurasi hyperparameter yang optimal guna meningkatkan akurasi prediksi, dan (3) mengidentifikasi fitur yang memiliki pengaruh signifikan dalam prediksi risiko asteroid. Algoritma Light Gradient Boosting Machine (LightGBM) menghasilkan kinerja model yang sangat baik dengan akurasi sebesar 99,97%, recall 93,37%, precision 92,08%, F1-score 92,72%, specificity 99,99%, false positive rate (FPR) 0,0139%, dan false negative rate (FNR) sebesar 6,63%. Kinerja optimal diperoleh melalui konfigurasi hyperparameter terbaik, yaitu num leaves = 10, learning rate = 0,1, dan n estimators = 15, yang merupakan kombinasi terbaik dari berbagai konfigurasi yang telah diuji. Selain itu, evaluasi kontribusi fitur menggunakan metode feature importance berbasis gain menunjukkan bahwa fitur argumen perihelion, magnitudo absolut, sumbu semi-mayor, dan bujur simpul naik merupakan fitur-fitur yang paling berpengaruh dalam pembentukan model, sementara fitur class asteroid tidak memberikan kontribusi karena tidak digunakan dalam proses pelatihan model.

Asteroid risk analysis has become a major concern in the field of astronomy, in line with the increasing volume of data generated by telescopes and satellites. In this study, the Light Gradient Boosting Machine (LightGBM) algorithm was implemented to classify asteroids based on their potential risk to Earth by utilizing multidimensional features. The dataset was obtained from the NASA JPL Small Body Database Query. This research aims to: (1) analyze the performance of LightGBM in identifying and predicting hazardous asteroids, (2) determine the optimal hyperparameter configuration to improve prediction accuracy, and (3) identify features that significantly influence asteroid risk prediction. The LightGBM algorithm yielded excellent model performance with an accuracy of 99.97%, a recall of 93.37%, precision of 92.08%, F1-score of 92.72%, specificity of 99.99%, false positive rate (FPR) of 0.0139%, and false negative rate (FNR) of 6.63%. Optimal performance was achieved using the best hyperparameter configuration, consisting of num leaves = 10, learning rate = 0.1, and n estimators = 15, which was the most effective combination among the various configurations tested. Furthermore, feature contribution evaluation using the gain-based feature importance method revealed that the argument of perihelion, absolute magnitude, semi-major axis, and longitude of ascending node were the most influential features in model construction, whereas the asteroid class feature contributed nothing, as it was excluded from the training process.