天體分類是天文學的基礎工作,因應天文觀測數據量的急速增加,以往的分類方式已無法有效率的處理待分類星體資料。本研究提出一套以數個監督式類神經網路模型共同組成之恆星分類流程,網路模型使用開源軟體庫Tensorflow的應用程式介面Keras建置,其具有視覺化及平行化計算優勢,並輔以其他相關Python套件作為分析工具。本研究以來自SDSS (Sloan Digital Sky Survey)、UKIDSS (UKIRT Infrared Deep Sky Survey)、WISE(Wide-field Infrared Survey Explorer)三個巡天觀測之星體光度資訊作為網路模型之訓練資料,以摩根-肯納(Morgan–Keenan classification)的七大類別及相關子類別作為訓練目標,透過調整參數設定和比較各類神經網路模型之表現,探討應用光度資訊作為訓練資料之可行性與分類結果的意義。研究結果顯示本分類方法於M、K類別達到98%與93%的正確率,已可上線投入實際應用。
Stellar classification is a fundamental work in astronomy. Since the classifier used today is overwhelmed by the rapidly growth of data set, new methods should be developed. In this study, A stellar classification is proposed, consists of several neural network models built by open source library Tensorflow, which is well-known by its visualize ability and supportive of computing ability to one or more CPUs or GPUs. 50043 stars are classified into Morgan–Keenan (MK) system using multi-band photometry extract from Sloan Digital Sky Survey (SDSS), UKIDSS, and WISE. The result demonstrates the feasibility of using multi-band photometry as input by comparing different setting of parameters and revealing the performance of every classifier. Models can reach 98% and 93% accuracy while classifying M type and K type and it's ready to serve online.