Image-Based Time-Series Representations and Transfer Learning for Enhanced Exoplanet Detection

Identifying planets outside our solar system orbiting distant stars is a difficult task in the field of astrophysics. Significant class imbalances are present in available light curve datasets, which adds complexity to this challenge. There are many more phenomena unrelated to exoplanets than there are actual exoplanets. This lack of balance affects the processing and classification of data, resulting in less than optimal performance based on evaluation metrics. This study explores how transfer learning (TL) can improve the identification of exoplanets. The core of this method is the transformation of light curves into formats based on images, using sophisticated signal processing techniques like the continuous wavelet transform (CWT) and Gramian Angular Field (GAF). These changes make it easier to uncover and recognize intricate patterns in the data that would be hard to distinguish in their original time-series structure. The modified datasets are later examined using advanced deep learning models such as MobileNet, ResNet-50, and Inception-v3. These models, well-known for their effectiveness in identifying images, are modified to suit the specific requirements of exoplanet identification. The MobileNet model shows encouraging results, with an 80% recall rate for the minority class based on empirical evidence. This represents a major development in accurately recognizing exoplanet occurrences in highly unbalanced datasets.