Spatio-temporal Deep Learning for Moving-object Tracklet Verification in Wide-field Optical Surveys
Abstract
Wide-field optical surveys generate enormous numbers of moving-object candidate detections, making reliable tracklet verification a critical component of modern survey pipelines. Existing real–bogus (R/B) classifiers typically evaluate individual detections based on their morphological and photometric properties, without fully exploiting the temporal continuity that characterizes genuine moving objects. In this work, two spatio-temporal deep-learning architectures for tracklet-level verification are proposed, including a lightweight convolutional neural network and bidirectional long short-term memory (CNN-BiLSTM) model and a vision transformer (ViT)-based sequence classifier. An extensive dataset acquired with a telescope array is used for model training, comprising 58,576 tracklets and approximately 611,000 image cutouts. Both models are trained and evaluated under identical protocols using fixed-length tracklet sequences together with auxiliary photometric and astrometric features. The results demonstrate that the CNN-BiLSTM consistently achieves a high recall rate for real correlated tracklets (>98%) while maintaining substantially higher precision (>90%) for uncorrelated tracklets than the ViT (<80%). In addition, it requires significantly less memory and delivers more than twice the inference throughput of the transformer-based model. Manual inspection of model–catalog discrepancy sets reveals that many apparent classification errors correspond to visually convincing moving objects that are absent from the reference catalog, indicating that the true predictive performance of the models is likely underestimated by catalog-based evaluation alone. These findings demonstrate that spatio-temporal deep learning not only improves operational tracklet verification but also serves as a complementary tool for identifying catalog incompleteness and supporting survey quality assessment in future wide-field optical surveys.